retrieval_evaluation.ipynb

{
 "cells": [
  {
   "cell_type": "code",
   "execution_count": 1,
   "metadata": {},
   "outputs": [],
   "source": [
    "\"\"\"\n",
    "Core dependencies for LLM evaluation framework\n",
    "Including:\n",
    "- Basic data processing and statistical libraries\n",
    "- Google Cloud authentication\n",
    "- LangChain components for retrieval and embeddings \n",
    "- Ragas evaluation metrics\n",
    "- PDF processing utilities\n",
    "\"\"\"\n",
    "\n",
    "# Standard libraries for data processing and analysis\n",
    "import os\n",
    "import time\n",
    "import numpy as np\n",
    "from scipy import stats\n",
    "from typing import List, Dict, Any\n",
    "from tqdm.notebook import tqdm\n",
    "import pandas as pd\n",
    "from os import stat\n",
    "import chromadb\n",
    "\n",
    "# Google Cloud authentication utilities\n",
    "from google.auth import load_credentials_from_file\n",
    "\n",
    "# LangChain components for retrieval and LLM integration\n",
    "from langchain_community.retrievers import BM25Retriever\n",
    "from langchain_core.documents import Document\n",
    "from langchain_google_vertexai import VertexAI, VertexAIEmbeddings\n",
    "from langchain_chroma import Chroma\n",
    "from langchain_openai import ChatOpenAI\n",
    "\n",
    "# Ragas evaluation metrics and utilities\n",
    "from ragas.metrics import LLMContextRecall, Faithfulness, FactualCorrectness, SemanticSimilarity\n",
    "from ragas import evaluate\n",
    "from ragas.llms import LangchainLLMWrapper\n",
    "from ragas.embeddings import LangchainEmbeddingsWrapper\n",
    "from ragas.run_config import RunConfig\n",
    "from ragas import EvaluationDataset\n",
    "\n",
    "# Type hints for better code clarity\n",
    "from typing import cast as t\n",
    "from langchain_core.outputs import LLMResult, ChatGeneration\n",
    "from langchain_core.messages import BaseMessage\n",
    "\n",
    "# Custom evaluation metrics\n",
    "from llm_evaluation.evaluate_fact_beta import FactualCorrectnessReviseBeta\n",
    "\n",
    "# PDF document processing\n",
    "import fitz  # PyMuPDF for PDF extraction"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 2,
   "metadata": {},
   "outputs": [],
   "source": [
    "# Load Google Cloud credentials and project configuration\n",
    "# Replace with your own credentials file\n",
    "credentials, project_id = load_credentials_from_file(\n",
    "    \"path/to/your/auth.json\"\n",
    ")\n",
    "os.environ[\"GOOGLE_APPLICATION_CREDENTIALS\"] = \"path/to/your/auth.json\""
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 1,
   "metadata": {},
   "outputs": [],
   "source": [
    "# Configure OpenAI API authentication\n",
    "import os \n",
    "os.environ[\"OPENAI_API_KEY\"] = \"<YOUR_OPENAI_API_KEY>\" "
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 4,
   "metadata": {},
   "outputs": [],
   "source": [
    "class CustomHybridRetriever:\n",
    "    \"\"\"\n",
    "    A hybrid retrieval system that combines vector-based and BM25 retrievers\n",
    "    \n",
    "    Implements a weighted combination of semantic search and keyword-based retrieval\n",
    "    to leverage the strengths of both approaches\n",
    "    \"\"\"\n",
    "    def __init__(self, vector_retriever, bm25_retriever, weight_vector=0.5, k=4):\n",
    "        \"\"\"\n",
    "        Initialize hybrid retriever with configurable weights and parameters\n",
    "        \n",
    "        Args:\n",
    "            vector_retriever: Semantic vector-based retriever\n",
    "            bm25_retriever: Keyword-based BM25 retriever\n",
    "            weight_vector: Weight for vector scores (0-1)\n",
    "            k: Number of documents to return\n",
    "        \"\"\"\n",
    "        self.vector_retriever = vector_retriever\n",
    "        self.bm25_retriever = bm25_retriever\n",
    "        self.weight_vector = weight_vector\n",
    "        self.k = k\n",
    "        \n",
    "    def invoke(self, query: str) -> List[Document]:\n",
    "        \"\"\"\n",
    "        Retrieve and combine results from both retrievers\n",
    "        \n",
    "        Process:\n",
    "        1. Get documents from both retrievers\n",
    "        2. Normalize relevance scores \n",
    "        3. Combine scores using weighted average\n",
    "        4. Return top-k documents\n",
    "        \n",
    "        Args:\n",
    "            query: Search query string\n",
    "            \n",
    "        Returns:\n",
    "            List of most relevant documents with combined scores\n",
    "        \"\"\"\n",
    "        try:\n",
    "            vector_docs = self.vector_retriever.invoke(query)\n",
    "            bm25_docs = self.bm25_retriever.invoke(query)\n",
    "            \n",
    "            print(f\"Vector retriever returned {len(vector_docs)} docs\")\n",
    "            print(f\"BM25 retriever returned {len(bm25_docs)} docs\")\n",
    "            \n",
    "            doc_scores = {}\n",
    "            \n",
    "            max_vector_score = max((doc.metadata.get('score', 0.0) for doc in vector_docs), default=1.0)\n",
    "            for doc in vector_docs:\n",
    "                content = doc.page_content\n",
    "                score = doc.metadata.get('score', 0.5)\n",
    "                # Normalize score\n",
    "                normalized_score = score / max_vector_score if max_vector_score > 0 else score\n",
    "                doc_scores[content] = {\n",
    "                    'vector_score': normalized_score,\n",
    "                    'bm25_score': 0.0,\n",
    "                    'doc': doc\n",
    "                }\n",
    "            \n",
    "            # Normalize scores for BM25 docs\n",
    "            max_bm25_score = max((doc.metadata.get('score', 0.0) for doc in bm25_docs), default=1.0)\n",
    "            for doc in bm25_docs:\n",
    "                content = doc.page_content\n",
    "                score = doc.metadata.get('score', 0.5)\n",
    "                normalized_score = score / max_bm25_score if max_bm25_score > 0 else score\n",
    "                \n",
    "                if content in doc_scores:\n",
    "                    doc_scores[content]['bm25_score'] = normalized_score\n",
    "                else:\n",
    "                    doc_scores[content] = {\n",
    "                        'vector_score': 0.0,\n",
    "                        'bm25_score': normalized_score,\n",
    "                        'doc': doc\n",
    "                    }\n",
    "            \n",
    "            # Calculate combined scores\n",
    "            final_scores = []\n",
    "            for content, scores in doc_scores.items():\n",
    "                combined_score = (\n",
    "                    self.weight_vector * scores['vector_score'] +\n",
    "                    (1 - self.weight_vector) * scores['bm25_score']\n",
    "                )\n",
    "                final_scores.append((combined_score, scores['doc']))\n",
    "            \n",
    "            # Sort by combined score and get top k\n",
    "            final_docs = [\n",
    "                doc for _, doc in sorted(final_scores, key=lambda x: x[0], reverse=True)[:self.k]\n",
    "            ]\n",
    "            \n",
    "            # Add combined scores to metadata\n",
    "            for i, doc in enumerate(final_docs):\n",
    "                doc.metadata['hybrid_score'] = final_scores[i][0]\n",
    "            \n",
    "            print(f\"Hybrid retriever returning {len(final_docs)} docs\")\n",
    "            if final_docs:\n",
    "                print(f\"Top document scores: {[doc.metadata.get('hybrid_score', 0.0) for doc in final_docs[:3]]}\")\n",
    "            \n",
    "            return final_docs\n",
    "            \n",
    "        except Exception as e:\n",
    "            print(f\"Error in hybrid retriever: {str(e)}\")\n",
    "            import traceback\n",
    "            traceback.print_exc()\n",
    "            return []"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 5,
   "metadata": {},
   "outputs": [],
   "source": [
    "def extract_text_from_pdf(pdf_path: str) -> List[Document]:\n",
    "    \"\"\"\n",
    "    Extract text content and metadata from PDF documents\n",
    "    \n",
    "    Process:\n",
    "    1. Opens PDF file using PyMuPDF\n",
    "    2. Extracts text from each page\n",
    "    3. Creates Document objects with content and metadata\n",
    "    \n",
    "    Args:\n",
    "        pdf_path: Path to the PDF file to process\n",
    "        \n",
    "    Returns:\n",
    "        List[Document]: Collection of Document objects containing:\n",
    "            - page_content: Extracted text from each page\n",
    "            - metadata: Source file, page number, and total pages\n",
    "    \"\"\"\n",
    "    documents = []\n",
    "    try:\n",
    "        # Open PDF document\n",
    "        pdf_document = fitz.open(pdf_path)\n",
    "\n",
    "        # Process each page\n",
    "        for page_num in range(len(pdf_document)):\n",
    "            page = pdf_document[page_num]\n",
    "            text = page.get_text(\"text\")\n",
    "\n",
    "            # Create document if page has content\n",
    "            if text.strip():\n",
    "                metadata = {\n",
    "                    \"source\": pdf_path,\n",
    "                    \"page\": page_num + 1,\n",
    "                    \"total_pages\": len(pdf_document)\n",
    "                }\n",
    "                documents.append(Document(page_content=text, metadata=metadata))\n",
    "        pdf_document.close()\n",
    "    except Exception as e:\n",
    "        print(f\"Error processing {pdf_path}: {str(e)}\")\n",
    "    return documents"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 6,
   "metadata": {},
   "outputs": [],
   "source": [
    "def load_pdfs_from_directory(directory_path: str) -> List[Document]:\n",
    "    \"\"\"\n",
    "    Load all PDF documents from a specified directory\n",
    "    \n",
    "    Args:\n",
    "        directory_path: Path to directory containing PDF files\n",
    "        \n",
    "    Returns:\n",
    "        List of Document objects from all PDFs in directory\n",
    "    \"\"\"\n",
    "    documents = []\n",
    "    with tqdm(os.listdir(directory_path), desc=\"Loading PDFs\") as pbar:\n",
    "        for filename in pbar:\n",
    "            if filename.lower().endswith('.pdf'):\n",
    "                file_path = os.path.join(directory_path, filename)\n",
    "                pbar.set_postfix(file=filename)\n",
    "                documents.extend(extract_text_from_pdf(file_path))\n",
    "    return documents"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 7,
   "metadata": {},
   "outputs": [],
   "source": [
    "def custom_is_finished_parser(response: LLMResult):\n",
    "    \"\"\"\n",
    "    Custom parser to determine if LLM generation is complete\n",
    "    \n",
    "    Checks completion status through:\n",
    "    1. Generation info metadata\n",
    "    2. Chat message response metadata\n",
    "    3. Default completion assumption\n",
    "    \n",
    "    Args:\n",
    "        response: LLM generation result object\n",
    "        \n",
    "    Returns:\n",
    "        bool: True if all generations are complete\n",
    "    \"\"\"\n",
    "    is_finished_list = []\n",
    "    for g in response.flatten():\n",
    "        resp = g.generations[0][0]\n",
    "\n",
    "        # Check generation info metadata\n",
    "        if resp.generation_info is not None:\n",
    "            if resp.generation_info.get(\"finish_reason\") is not None:\n",
    "                is_finished_list.append(\n",
    "                    resp.generation_info.get(\"finish_reason\") == \"STOP\"\n",
    "                )\n",
    "\n",
    "        # Check chat message response metadata\n",
    "        elif (\n",
    "            isinstance(resp, ChatGeneration)\n",
    "            and t.cast(ChatGeneration, resp).message is not None\n",
    "        ):\n",
    "            resp_message: BaseMessage = t.cast(ChatGeneration, resp).message\n",
    "            if resp_message.response_metadata.get(\"finish_reason\") is not None:\n",
    "                is_finished_list.append(\n",
    "                    resp_message.response_metadata.get(\"finish_reason\") == \"STOP\"\n",
    "                )\n",
    "\n",
    "        # Default to complete if no metadata available\n",
    "        else:\n",
    "            is_finished_list.append(True)\n",
    "    return all(is_finished_list)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 8,
   "metadata": {},
   "outputs": [],
   "source": [
    "class RetrievalEvaluator:\n",
    "    \"\"\"\n",
    "    Evaluation framework for comparing different retrieval methods\n",
    "    \n",
    "    Implements comprehensive evaluation of vector, BM25, and hybrid retrieval approaches\n",
    "    using multiple metrics and statistical analysis\n",
    "    \"\"\"\n",
    "    def __init__(self, data_dir: str, persist_dir: str):\n",
    "        \"\"\"\n",
    "        Initialize evaluation framework with models and retrievers\n",
    "        \n",
    "        Args:\n",
    "            data_dir: Directory containing source documents\n",
    "            persist_dir: Directory for storing vector indices\n",
    "        \"\"\"\n",
    "        self.data_dir = data_dir\n",
    "        self.persist_dir = persist_dir\n",
    "        \n",
    "        # Initialize embeddings\n",
    "        self.embeddings = VertexAIEmbeddings(\n",
    "            model_name=\"textembedding-gecko\",\n",
    "            project=project_id,\n",
    "            credentials=credentials,\n",
    "            location=\"us-central1\"\n",
    "        )\n",
    "        \n",
    "        # Initialize LLMs\n",
    "        self.eval_llm = ChatOpenAI(\n",
    "            model=\"gpt-3.5-turbo\",\n",
    "            temperature=0.1\n",
    "        )\n",
    "        \n",
    "        self.exp_llm = VertexAI(\n",
    "            model_name=\"gemini-1.0-pro\",\n",
    "            temperature=0.1,\n",
    "            max_output_tokens=8192,\n",
    "            project=project_id,\n",
    "            credentials=credentials\n",
    "        )\n",
    "        \n",
    "        print(\"Loading documents...\")\n",
    "        self.documents = load_pdfs_from_directory(data_dir)\n",
    "        print(f\"Loaded {len(self.documents)} documents\")\n",
    "        \n",
    "        print(\"Initializing retrievers...\")\n",
    "        self.vector_retriever = self._init_vector_retriever()\n",
    "        self.bm25_retriever = self._init_bm25_retriever()\n",
    "        self.hybrid_retriever = self._init_hybrid_retriever()\n",
    "        print(\"Initialization complete\")\n",
    "\n",
    "    def _init_vector_retriever(self):\n",
    "        \"\"\"\n",
    "        Initialize vector-based retrieval system\n",
    "        \n",
    "        Process:\n",
    "        1. Clear existing vector store\n",
    "        2. Filter valid documents\n",
    "        3. Create new vector embeddings\n",
    "        4. Configure retriever parameters\n",
    "        \n",
    "        Returns:\n",
    "            Configured vector retriever or None if initialization fails\n",
    "        \"\"\"\n",
    "        print(\"Initializing vector retriever...\")\n",
    "        try:\n",
    "            if os.path.exists(f\"{self.persist_dir}/vector\"):\n",
    "                import shutil\n",
    "                shutil.rmtree(f\"{self.persist_dir}/vector\")\n",
    "                print(\"Cleared existing vector store\")\n",
    "\n",
    "            print(\"Creating new vector store...\")\n",
    "\n",
    "            valid_docs = [doc for doc in self.documents if doc.page_content.strip()]\n",
    "            print(f\"Found {len(valid_docs)} valid documents\")\n",
    "\n",
    "            vectorstore = Chroma.from_documents(\n",
    "                documents=valid_docs,\n",
    "                embedding=self.embeddings,\n",
    "                persist_directory=f\"{self.persist_dir}/vector\"\n",
    "            )\n",
    "            print(f\"Created vector store with {vectorstore._collection.count()} documents\")\n",
    "\n",
    "            retriever = vectorstore.as_retriever(\n",
    "                search_kwargs={\n",
    "                    \"k\": 4,\n",
    "                }\n",
    "            )\n",
    "            \n",
    "            return retriever\n",
    "            \n",
    "        except Exception as e:\n",
    "            print(f\"Error initializing vector retriever: {str(e)}\")\n",
    "            import traceback\n",
    "            traceback.print_exc()\n",
    "            return None\n",
    "\n",
    "    def _init_bm25_retriever(self):\n",
    "        \"\"\"\n",
    "        Initialize BM25 keyword-based retriever\n",
    "        \n",
    "        Process:\n",
    "        1. Preprocess documents\n",
    "        2. Configure BM25 parameters\n",
    "        3. Validate retriever functionality\n",
    "        \n",
    "        Returns:\n",
    "            Configured BM25 retriever or None if initialization fails\n",
    "        \"\"\"\n",
    "        print(\"Initializing BM25 retriever...\")\n",
    "        try:\n",
    "            processed_docs = []\n",
    "            for doc in self.documents:\n",
    "                if doc.page_content.strip():  \n",
    "                    metadata = doc.metadata.copy() if hasattr(doc, 'metadata') else {}\n",
    "                    processed_docs.append(Document(\n",
    "                        page_content=doc.page_content,\n",
    "                        metadata=metadata\n",
    "                    ))\n",
    "            \n",
    "            print(f\"Processing {len(processed_docs)} documents for BM25\")\n",
    "            \n",
    "\n",
    "            retriever = BM25Retriever.from_documents(\n",
    "                documents=processed_docs,\n",
    "                preprocess_func=lambda text: text.split(), \n",
    "                k=4\n",
    "            )\n",
    "            \n",
    "            test_query = \"What is instruction tuning?\"\n",
    "            results = retriever.invoke(test_query)\n",
    "            print(f\"Test query returned {len(results)} documents\")\n",
    "            if results:\n",
    "                print(\"First document preview:\", results[0].page_content[:200])\n",
    "            \n",
    "            return retriever\n",
    "            \n",
    "        except Exception as e:\n",
    "            print(f\"Error initializing BM25 retriever: {str(e)}\")\n",
    "            import traceback\n",
    "            traceback.print_exc()\n",
    "            return None\n",
    "\n",
    "    def _init_hybrid_retriever(self):\n",
    "        \"\"\"Initialize hybrid retriever with optimized weights\"\"\"\n",
    "        print(\"Initializing hybrid retriever...\")\n",
    "        return CustomHybridRetriever(\n",
    "            self.vector_retriever,\n",
    "            self.bm25_retriever,\n",
    "            weight_vector=0.3,\n",
    "            k=4\n",
    "        )\n",
    "    \n",
    "    def get_vector_retriever(self):\n",
    "        \"\"\"Get vector retriever\"\"\"\n",
    "        return self.vector_retriever\n",
    "    \n",
    "    def get_bm25_retriever(self):\n",
    "        \"\"\"Get BM25 retriever\"\"\"\n",
    "        return self.bm25_retriever\n",
    "    \n",
    "    def get_hybrid_retriever(self):\n",
    "        \"\"\"Get hybrid retriever\"\"\"\n",
    "        return self.hybrid_retriever\n",
    "\n",
    "    def _generate_response(self, query: str, retrieved_docs: List[Document]) -> str:\n",
    "        \"\"\"Generate answer based on retrieved documents\"\"\"\n",
    "        try:\n",
    "            valid_docs = []\n",
    "            for doc in retrieved_docs:\n",
    "                if isinstance(doc, Document) and doc.page_content.strip():\n",
    "                    valid_docs.append(doc)\n",
    "            \n",
    "            if not valid_docs:\n",
    "                print(\"Warning: No valid documents for response generation\")\n",
    "                return \"\"\n",
    "            \n",
    "            context = \"\\n\\n\".join([doc.page_content for doc in valid_docs])\n",
    "            prompt = f\"Based on the following context, please answer the question:\\n\\nContext: {context}\\n\\nQuestion: {query}\"\n",
    "            \n",
    "            response = self.exp_llm.invoke(prompt)\n",
    "            return response.content if hasattr(response, 'content') else str(response)\n",
    "            \n",
    "        except Exception as e:\n",
    "            print(f\"Error generating response: {str(e)}\")\n",
    "            return \"\"\n",
    "        \n",
    "    def _process_query(self, retriever, query: str) -> List[Document]:\n",
    "        \"\"\"Process a single query with improved logging\"\"\"\n",
    "        try:\n",
    "            print(f\"\\nProcessing query: {query[:100]}...\")\n",
    "            \n",
    "            retrieved_docs = retriever.invoke(query)\n",
    "            \n",
    "            print(f\"Retrieved {len(retrieved_docs)} documents\")\n",
    "            if retrieved_docs:\n",
    "                for i, doc in enumerate(retrieved_docs):\n",
    "                    print(f\"\\nDocument {i+1}:\")\n",
    "                    print(f\"Length: {len(doc.page_content)}\")\n",
    "                    print(f\"Content preview: {doc.page_content[:200]}...\")\n",
    "                    if hasattr(doc, 'metadata'):\n",
    "                        print(f\"Metadata: {doc.metadata}\")\n",
    "                    if hasattr(doc, 'similarity'):\n",
    "                        print(f\"Similarity score: {doc.similarity}\")\n",
    "            else:\n",
    "                print(\"Warning: No documents retrieved\")\n",
    "                \n",
    "            return retrieved_docs\n",
    "            \n",
    "        except Exception as e:\n",
    "            print(f\"Error processing query: {str(e)}\")\n",
    "            return []\n",
    "\n",
    "\n",
    "    def evaluate_retriever(self, retriever_name: str, test_dataset: pd.DataFrame, n_runs: int = 3):\n",
    "        \"\"\"\n",
    "        Evaluate retriever performance with statistical analysis\n",
    "        \n",
    "        Process:\n",
    "        1. Run multiple evaluation iterations\n",
    "        2. Calculate metrics for each run\n",
    "        3. Compute statistical measures (mean, confidence intervals)\n",
    "        4. Generate detailed performance report\n",
    "        \n",
    "        Args:\n",
    "            retriever_name: Type of retriever to evaluate\n",
    "            test_dataset: Evaluation queries and ground truth\n",
    "            n_runs: Number of evaluation iterations\n",
    "            \n",
    "        Returns:\n",
    "            Dict containing evaluation results and statistics\n",
    "        \"\"\"\n",
    "        try:\n",
    "            # Get retriever\n",
    "            if retriever_name == \"hybrid\":\n",
    "                retriever = self.get_hybrid_retriever()\n",
    "            elif retriever_name == \"vector\":\n",
    "                retriever = self.get_vector_retriever()\n",
    "            elif retriever_name == \"bm25\":\n",
    "                retriever = self.get_bm25_retriever()\n",
    "            else:\n",
    "                raise ValueError(f\"Unknown retriever type: {retriever_name}\")\n",
    "                \n",
    "            if retriever is None:\n",
    "                print(f\"Failed to initialize {retriever_name} retriever\")\n",
    "                return None\n",
    "            \n",
    "            # Store results from multiple runs\n",
    "            all_run_scores = []\n",
    "            \n",
    "            print(f\"\\nRunning {n_runs} evaluations for {retriever_name}...\")\n",
    "            \n",
    "            for run in range(n_runs):\n",
    "                print(f\"\\nStarting run {run + 1}/{n_runs}\")\n",
    "                \n",
    "                # Create evaluation dataset\n",
    "                valid_results = pd.DataFrame()\n",
    "                valid_results['user_input'] = test_dataset['question']\n",
    "                valid_results['reference'] = test_dataset['ground_truth']\n",
    "                valid_results['response'] = ''\n",
    "                valid_results['retrieved_contexts'] = [[]] * len(test_dataset)\n",
    "\n",
    "                print(f\"Processing {len(valid_results)} valid queries\")\n",
    "                \n",
    "                # Process each query\n",
    "                for idx, row in valid_results.iterrows():\n",
    "                    try:\n",
    "                        query = row['user_input']\n",
    "                        if pd.isna(query):\n",
    "                            continue\n",
    "                            \n",
    "                        # Retrieve documents with detailed logging\n",
    "                        print(f\"\\nProcessing query {idx + 1}/{len(valid_results)}\")\n",
    "                        print(f\"Query: {query[:100]}...\")\n",
    "                        \n",
    "                        if retriever_name == \"vector\":\n",
    "                            # For vector retriever, use get_relevant_documents\n",
    "                            retrieved_docs = retriever.get_relevant_documents(query)\n",
    "                        else:\n",
    "                            # For other retrievers, use invoke\n",
    "                            retrieved_docs = retriever.invoke(query)\n",
    "                            \n",
    "                        if retriever_name == \"bm25\":\n",
    "                            retrieved_docs = retriever.invoke(query)\n",
    "                            retrieved_docs = [\n",
    "                                Document(\n",
    "                                    page_content=doc.page_content,\n",
    "                                    metadata=doc.metadata if hasattr(doc, 'metadata') else {}\n",
    "                                ) for doc in retrieved_docs\n",
    "                            ]\n",
    "                        else:\n",
    "                            retrieved_docs = retriever.invoke(query)\n",
    "                        \n",
    "                        print(f\"Retrieved {len(retrieved_docs)} documents\")\n",
    "                        \n",
    "                        if not retrieved_docs:\n",
    "                            print(f\"Warning: No documents retrieved for query: {query[:50]}...\")\n",
    "                            continue\n",
    "                        \n",
    "                        # Print first document details for debugging\n",
    "                        if retrieved_docs:\n",
    "                            print(\"First retrieved document:\")\n",
    "                            print(f\"Length: {len(retrieved_docs[0].page_content)}\")\n",
    "                            print(f\"Content preview: {retrieved_docs[0].page_content[:200]}...\")\n",
    "                            if hasattr(retrieved_docs[0], 'metadata'):\n",
    "                                print(f\"Metadata: {retrieved_docs[0].metadata}\")\n",
    "                        \n",
    "                        # Generate response\n",
    "                        response = self._generate_response(query, retrieved_docs)\n",
    "                        print(f\"Generated response length: {len(response)}\")\n",
    "                        \n",
    "                        # Update results\n",
    "                        valid_results.at[idx, 'response'] = response\n",
    "                        valid_results.at[idx, 'retrieved_contexts'] = [doc.page_content for doc in retrieved_docs]\n",
    "                        \n",
    "                        time.sleep(1)  # Avoid rate limits\n",
    "                        \n",
    "                    except Exception as e:\n",
    "                        print(f\"Error processing query: {query[:50]}...\")\n",
    "                        print(f\"Error: {str(e)}\")\n",
    "                        import traceback\n",
    "                        traceback.print_exc()\n",
    "                        continue\n",
    "\n",
    "                print(f\"Successfully processed {len(valid_results)} queries\")\n",
    "\n",
    "                # Create evaluation dataset\n",
    "                eval_dataset = EvaluationDataset.from_pandas(valid_results)\n",
    "\n",
    "                # Setup evaluators\n",
    "                evaluator_llm = LangchainLLMWrapper(self.eval_llm)\n",
    "                evaluator_embeddings = LangchainEmbeddingsWrapper(self.embeddings)\n",
    "\n",
    "                # Define metrics\n",
    "                metrics = [\n",
    "                    LLMContextRecall(llm=evaluator_llm),\n",
    "                    FactualCorrectnessReviseBeta(llm=evaluator_llm),\n",
    "                    Faithfulness(llm=evaluator_llm),\n",
    "                    SemanticSimilarity(embeddings=evaluator_embeddings)\n",
    "                ]\n",
    "\n",
    "                # Run evaluation\n",
    "                print(\"\\nRunning evaluation metrics...\")\n",
    "                results = evaluate(\n",
    "                    dataset=eval_dataset,\n",
    "                    metrics=metrics\n",
    "                )\n",
    "\n",
    "                # Process scores for this run\n",
    "                run_scores = {}\n",
    "                if hasattr(results, 'scores'):\n",
    "                    if isinstance(results.scores, list) and results.scores:\n",
    "                        first_result = results.scores[0]\n",
    "                        for metric_name, score in first_result.items():\n",
    "                            if isinstance(score, (int, float)) and not pd.isna(score):\n",
    "                                run_scores[metric_name] = float(score)\n",
    "                \n",
    "                if run_scores:\n",
    "                    all_run_scores.append(run_scores)\n",
    "                    print(f\"Run {run + 1} scores:\", run_scores)\n",
    "                else:\n",
    "                    print(f\"Warning: No valid scores for run {run + 1}\")\n",
    "            \n",
    "            # Calculate average scores and confidence intervals\n",
    "            if not all_run_scores:\n",
    "                print(f\"No valid results for {retriever_name}\")\n",
    "                return None\n",
    "                \n",
    "            # Calculate mean scores across all runs\n",
    "            mean_scores = {}\n",
    "            confidence_intervals = {}\n",
    "            \n",
    "            for metric in all_run_scores[0].keys():\n",
    "                metric_values = [run[metric] for run in all_run_scores if metric in run]\n",
    "                if metric_values:\n",
    "                    mean_score = np.mean(metric_values)\n",
    "                    std_dev = np.std(metric_values) if len(metric_values) > 1 else 0.1\n",
    "                    \n",
    "                    mean_scores[metric] = mean_score\n",
    "                    confidence_intervals[metric] = (\n",
    "                        max(0.0, mean_score - std_dev),  # Lower bound\n",
    "                        min(1.0, mean_score + std_dev)   # Upper bound\n",
    "                    )\n",
    "            \n",
    "            print(f\"\\nFinal average scores for {retriever_name}:\")\n",
    "            for metric, score in mean_scores.items():\n",
    "                ci = confidence_intervals[metric]\n",
    "                print(f\"{metric}: {score:.4f} (95% CI: [{ci[0]:.4f}, {ci[1]:.4f}])\")\n",
    "            \n",
    "            return {\n",
    "                \"mean_scores\": mean_scores,\n",
    "                \"confidence_intervals\": confidence_intervals,\n",
    "                \"all_run_scores\": all_run_scores,\n",
    "                \"raw_results\": results\n",
    "            }\n",
    "\n",
    "        except Exception as e:\n",
    "            print(f\"Error evaluating {retriever_name}: {str(e)}\")\n",
    "            import traceback\n",
    "            traceback.print_exc()\n",
    "            return None"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 9,
   "metadata": {},
   "outputs": [],
   "source": [
    "def visualize_t_validation(results, save_path=None):\n",
    "    \"\"\"\n",
    "    Create individual T-Validation comparison plots for each evaluation metric\n",
    "    \n",
    "    Process:\n",
    "    1. Generates separate plots for each metric\n",
    "    2. Calculates error bars from confidence intervals\n",
    "    3. Formats plots with grid, labels, and titles\n",
    "    4. Optionally saves plots to specified directory\n",
    "    \n",
    "    Args:\n",
    "        results: Dictionary containing evaluation results and statistics\n",
    "        save_path: Optional path to save generated plots\n",
    "    \"\"\"\n",
    "    import matplotlib.pyplot as plt\n",
    "    import numpy as np\n",
    "    \n",
    "    # Define evaluation metrics to visualize\n",
    "    metrics = [\n",
    "        'context_recall',\n",
    "        'factual_correctness', \n",
    "        'faithfulness',\n",
    "        'semantic_similarity'\n",
    "    ]\n",
    "    \n",
    "    # Create separate plot for each metric\n",
    "    for metric in metrics:\n",
    "        plt.figure(figsize=(10, 6))\n",
    "        \n",
    "        retrievers = []\n",
    "        means = []\n",
    "        errors = []\n",
    "\n",
    "        # Collect data for visualization\n",
    "        for retriever_name, result in results.items():\n",
    "            if result and result.get(\"mean_scores\") and metric in result[\"mean_scores\"]:\n",
    "                score = result[\"mean_scores\"][metric]\n",
    "                ci = result[\"confidence_intervals\"][metric]\n",
    "                error = (ci[1] - ci[0]) / 2  # Calculate error range from CI\n",
    "                \n",
    "                retrievers.append(retriever_name)\n",
    "                means.append(score)\n",
    "                errors.append(error)\n",
    "        \n",
    "        if not retrievers:\n",
    "            print(f\"No valid results for metric: {metric}\")\n",
    "            continue\n",
    "            \n",
    "        x_pos = np.arange(len(retrievers))\n",
    "        \n",
    "        plt.errorbar(x_pos, means, \n",
    "                    yerr=errors,\n",
    "                    fmt='o', capsize=5, capthick=1.5, \n",
    "                    markersize=8, elinewidth=1.5)\n",
    "        \n",
    "        for i, mean in enumerate(means):\n",
    "            plt.text(x_pos[i], mean + errors[i], f'{mean:.3f}', \n",
    "                    horizontalalignment='center', verticalalignment='bottom')\n",
    "        \n",
    "        metric_name = metric.replace('_', ' ').title()\n",
    "        plt.title(f'T-Validation Comparison - {metric_name}')\n",
    "        plt.xlabel('Retriever Type')\n",
    "        plt.ylabel(f'{metric_name} Score')\n",
    "        plt.xticks(x_pos, retrievers, rotation=45)\n",
    "\n",
    "        plt.grid(True, linestyle='--', alpha=0.7)\n",
    "        plt.ylim(0, 1)\n",
    "        plt.tight_layout()\n",
    "\n",
    "        if save_path:\n",
    "            os.makedirs(save_path, exist_ok=True)\n",
    "            plt.savefig(f'{save_path}/t_validation_{metric}.png', \n",
    "                       dpi=300, bbox_inches='tight')\n",
    "        \n",
    "        plt.show()"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 10,
   "metadata": {},
   "outputs": [],
   "source": [
    "def run_evaluation(data_dir: str = \"./data\", persist_dir: str = \"./chroma_db\", n_samples: int = 15, n_runs: int = 3):\n",
    "    \"\"\"\n",
    "    Execute comprehensive evaluation of retrieval systems\n",
    "    \n",
    "    Process:\n",
    "    1. Initialize evaluation framework\n",
    "    2. Load and preprocess test dataset\n",
    "    3. Evaluate each retriever type\n",
    "    4. Generate statistical analysis\n",
    "    5. Create visualization plots\n",
    "    \n",
    "    Args:\n",
    "        data_dir: Directory containing source documents\n",
    "        persist_dir: Directory for vector store\n",
    "        n_samples: Number of test samples to evaluate\n",
    "        n_runs: Number of evaluation iterations\n",
    "        \n",
    "    Returns:\n",
    "        Dictionary containing evaluation results for all retrievers\n",
    "    \"\"\"\n",
    "    print(\"Starting evaluation process...\")\n",
    "    \n",
    "    # Initialize evaluation framework\n",
    "    evaluator = RetrievalEvaluator(data_dir, persist_dir)\n",
    "    \n",
    "    # Load and prepare test dataset\n",
    "    print(\"Loading test dataset...\")\n",
    "    test_dataset = pd.read_json('./data/eval_dataset_1.json')\n",
    "    if n_samples:\n",
    "        test_dataset = test_dataset.head(n_samples)\n",
    "    \n",
    "    # Evaluate each retriever type\n",
    "    results = {}\n",
    "    retrievers = [\"vector\", \"bm25\", \"hybrid\"]\n",
    "    \n",
    "    for retriever in tqdm(retrievers, desc=\"Evaluating retrievers\"):\n",
    "        print(f\"\\nEvaluating {retriever}...\")\n",
    "        result = evaluator.evaluate_retriever(retriever, test_dataset, n_runs=n_runs)\n",
    "        if result and result.get(\"mean_scores\"):\n",
    "            results[retriever] = result\n",
    "    \n",
    "    # Display evaluation results\n",
    "    print(\"\\nEvaluation Results:\")\n",
    "    print(\"=\" * 50)\n",
    "    \n",
    "    for retriever_name in retrievers:\n",
    "        print(f\"\\nResults for {retriever_name}:\")\n",
    "        if retriever_name in results and results[retriever_name].get(\"mean_scores\"):\n",
    "            print(\"\\nMean scores across all runs:\")\n",
    "            for metric, score in results[retriever_name][\"mean_scores\"].items():\n",
    "                ci = results[retriever_name][\"confidence_intervals\"][metric]\n",
    "                print(f\"{metric}: {score:.4f} (95% CI: [{ci[0]:.4f}, {ci[1]:.4f}])\")\n",
    "        else:\n",
    "            print(\"Mean scores: None\")\n",
    "        print(\"-\" * 30)\n",
    "    \n",
    "    # Generate visualization plots\n",
    "    visualize_t_validation(results, save_path=\"./evaluation_plots\")\n",
    "    \n",
    "    return results"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 11,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Starting evaluation process...\n",
      "Loading documents...\n"
     ]
    },
    {
     "data": {
      "application/vnd.jupyter.widget-view+json": {
       "model_id": "be0e36b5d82644c8a35076fbe9215b63",
       "version_major": 2,
       "version_minor": 0
      },
      "text/plain": [
       "Loading PDFs:   0%|          | 0/11 [00:00<?, ?it/s]"
      ]
     },
     "metadata": {},
     "output_type": "display_data"
    },
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Loaded 290 documents\n",
      "Initializing retrievers...\n",
      "Initializing vector retriever...\n",
      "Cleared existing vector store\n",
      "Creating new vector store...\n",
      "Found 290 valid documents\n",
      "Created vector store with 290 documents\n",
      "Initializing BM25 retriever...\n",
      "Processing 290 documents for BM25\n",
      "Test query returned 4 documents\n",
      "First document preview: Published as a conference paper at ICLR 2022\n",
      "Nonetheless, the fact that scale does not always improve zero-shot performance of untuned LaMDA-\n",
      "PT is an interesting artifact. Initially, we were surprise\n",
      "Initializing hybrid retriever...\n",
      "Initialization complete\n",
      "Loading test dataset...\n"
     ]
    },
    {
     "data": {
      "application/vnd.jupyter.widget-view+json": {
       "model_id": "905223fc45c243019f8a8405848eecc2",
       "version_major": 2,
       "version_minor": 0
      },
      "text/plain": [
       "Evaluating retrievers:   0%|          | 0/3 [00:00<?, ?it/s]"
      ]
     },
     "metadata": {},
     "output_type": "display_data"
    },
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "\n",
      "Evaluating vector...\n",
      "\n",
      "Running 3 evaluations for vector...\n",
      "\n",
      "Starting run 1/3\n",
      "Processing 15 valid queries\n",
      "\n",
      "Processing query 1/15\n",
      "Query: How does instruction tuning affect the zero-shot performance on unseen tasks?...\n"
     ]
    },
    {
     "name": "stderr",
     "output_type": "stream",
     "text": [
      "/var/folders/tk/1ns3nr911378ctmw70qjs1700000gn/T/ipykernel_66589/3274679769.py:247: LangChainDeprecationWarning: The method `BaseRetriever.get_relevant_documents` was deprecated in langchain-core 0.1.46 and will be removed in 1.0. Use :meth:`~invoke` instead.\n",
      "  retrieved_docs = retriever.get_relevant_documents(query)\n"
     ]
    },
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Retrieved 4 documents\n",
      "First retrieved document:\n",
      "Length: 3976\n",
      "Content preview: Published as a conference paper at ICLR 2022\n",
      "4.2\n",
      "SCALING LAWS\n",
      "0.4B\n",
      "2B\n",
      "8B\n",
      "68B\n",
      "137B\n",
      "Model Size (# parameters)\n",
      "30\n",
      "40\n",
      "50\n",
      "60\n",
      "70\n",
      "Instruction tuning\n",
      "Untuned model\n",
      "Average zero-shot accuracy \n",
      "on 13 held-out t...\n",
      "Metadata: {'page': 7, 'source': './data/2109.01652.pdf', 'total_pages': 46}\n",
      "Generated response length: 2768\n",
      "\n",
      "Processing query 2/15\n",
      "Query: What is the Zero-shot-CoT method and how does it elicit chain of thought from large language models?...\n",
      "Retrieved 4 documents\n",
      "First retrieved document:\n",
      "Length: 3166\n",
      "Content preview: Large Language Models are Zero-Shot Reasoners\n",
      "Takeshi Kojima\n",
      "The University of Tokyo\n",
      "t.kojima@weblab.t.u-tokyo.ac.jp\n",
      "Shixiang Shane Gu\n",
      "Google Research, Brain Team\n",
      "Machel Reid\n",
      "Google Research∗\n",
      "Yutaka M...\n",
      "Metadata: {'page': 1, 'source': './data/2205.11916.pdf', 'total_pages': 42}\n",
      "Generated response length: 2635\n",
      "\n",
      "Processing query 3/15\n",
      "Query: How does prompt tuning affect model performance in improving NLP tasks?...\n",
      "Retrieved 4 documents\n",
      "First retrieved document:\n",
      "Length: 3925\n",
      "Content preview: Published as a conference paper at ICLR 2022\n",
      "instruct(xk)⊕yk ⊕instruct(x)”, where ⊕denotes string concatenation with a delimiter token inserted\n",
      "in between. At both training and inference time, exempla...\n",
      "Metadata: {'page': 8, 'source': './data/2109.01652.pdf', 'total_pages': 46}\n",
      "Generated response length: 2340\n",
      "\n",
      "Processing query 4/15\n",
      "Query: What is the purpose of instruction tuning in language models and how does it enhance response to NLP...\n",
      "Retrieved 4 documents\n",
      "First retrieved document:\n",
      "Length: 4454\n",
      "Content preview: Published as a conference paper at ICLR 2022\n",
      "et al., 2021). Compared with this prior work that ﬁnetunes and evaluates on the same downstream\n",
      "task, our setup studies the effect of instruction tuning on...\n",
      "Metadata: {'page': 29, 'source': './data/2109.01652.pdf', 'total_pages': 46}\n",
      "Generated response length: 2323\n",
      "\n",
      "Processing query 5/15\n",
      "Query: What distinguishes Zero-shot-CoT from Few-shot-CoT in terms of prompting methods for large language ...\n",
      "Retrieved 4 documents\n",
      "First retrieved document:\n",
      "Length: 3436\n",
      "Content preview: Table 4: Robustness study against template measured on the MultiArith dataset with text-davinci-002.\n",
      "(*1) This template is used in Ahn et al. [2022] where a language model is prompted to generate\n",
      "step...\n",
      "Metadata: {'page': 8, 'source': './data/2205.11916.pdf', 'total_pages': 42}\n",
      "Generated response length: 2145\n",
      "\n",
      "Processing query 6/15\n",
      "Query: Which language models were used in the experiment 'Exploring Zero-Shot Learning in Neural Networks',...\n",
      "Retrieved 4 documents\n",
      "First retrieved document:\n",
      "Length: 3052\n",
      "Content preview: Table 8: Description of language models. (*1) As for Original GPT3 models, we assign model size in-\n",
      "formation to each model by referring to https://blog.eleuther.ai/gpt3-model-sizes/ and\n",
      "https://beta....\n",
      "Metadata: {'page': 17, 'source': './data/2205.11916.pdf', 'total_pages': 42}\n",
      "Generated response length: 1908\n",
      "\n",
      "Processing query 7/15\n",
      "Query: How does Zero-shot-CoT differ from previous few-shot approaches in eliciting chain of thought from l...\n",
      "Retrieved 4 documents\n",
      "First retrieved document:\n",
      "Length: 4416\n",
      "Content preview: model (text-davinci-002). We also evaluate Zero-shot-CoT with another off-the-shelf large model,\n",
      "540B parameter PaLM, showing similar magnitudes of improvements on MultiArith and GSM8K.\n",
      "Importantly, w...\n",
      "Metadata: {'page': 3, 'source': './data/2205.11916.pdf', 'total_pages': 42}\n",
      "Generated response length: 2181\n",
      "\n",
      "Processing query 8/15\n",
      "Query: What are the stages in the Zero-shot-CoT method for reasoning and answer extraction, and how are the...\n",
      "Retrieved 4 documents\n",
      "First retrieved document:\n",
      "Length: 2726\n",
      "Content preview: Table 10: Answer extraction prompts used for Zero-shot-CoT experiments in Table 1. C.S.QA :\n",
      "CommonsenseQA, D.U. : Date Understanding, S.O. : Tracking Shufﬂed Objects\n",
      "No Task\n",
      "Zero-Shot-CoT (left)\n",
      "Zero-...\n",
      "Metadata: {'page': 18, 'source': './data/2205.11916.pdf', 'total_pages': 42}\n",
      "Generated response length: 2405\n",
      "\n",
      "Processing query 9/15\n",
      "Query: What are the main approaches for inducing LLMs to perform chain-of-thought reasoning, according to W...\n",
      "Retrieved 4 documents\n",
      "First retrieved document:\n",
      "Length: 4229\n",
      "Content preview: C\n",
      "Extended Related Work\n",
      "Chain-of-thought prompting is a general approach that is inspired by several prior directions: prompt-\n",
      "ing, natural language explanations, program synthesis/execution, numeric ...\n",
      "Metadata: {'page': 24, 'source': './data/2201.11903.pdf', 'total_pages': 43}\n",
      "Generated response length: 1827\n",
      "\n",
      "Processing query 10/15\n",
      "Query: Which sorting method has the most impact on Auto-CoT's demonstration accuracy in question clustering...\n",
      "Retrieved 4 documents\n",
      "First retrieved document:\n",
      "Length: 4087\n",
      "Content preview: the effect of misleading by similarity (Section 3.1). On the other hand, if we took each demonstration as a kind of skill,\n",
      "diverse demonstrations seem to cover more alternative skills for solving targ...\n",
      "Metadata: {'page': 6, 'source': './data/2210.03493.pdf', 'total_pages': 25}\n",
      "Generated response length: 2701\n",
      "\n",
      "Processing query 11/15\n",
      "Query: What are the pros and cons of prompting methods for large language models in terms of their cognitiv...\n",
      "Retrieved 4 documents\n",
      "First retrieved document:\n",
      "Length: 3166\n",
      "Content preview: Large Language Models are Zero-Shot Reasoners\n",
      "Takeshi Kojima\n",
      "The University of Tokyo\n",
      "t.kojima@weblab.t.u-tokyo.ac.jp\n",
      "Shixiang Shane Gu\n",
      "Google Research, Brain Team\n",
      "Machel Reid\n",
      "Google Research∗\n",
      "Yutaka M...\n",
      "Metadata: {'page': 1, 'source': './data/2205.11916.pdf', 'total_pages': 42}\n",
      "Generated response length: 2815\n",
      "\n",
      "Processing query 12/15\n",
      "Query: What are the stages in Zero-shot-CoT for reasoning and answer extraction, and how is it different fr...\n",
      "Retrieved 4 documents\n",
      "First retrieved document:\n",
      "Length: 2726\n",
      "Content preview: Table 10: Answer extraction prompts used for Zero-shot-CoT experiments in Table 1. C.S.QA :\n",
      "CommonsenseQA, D.U. : Date Understanding, S.O. : Tracking Shufﬂed Objects\n",
      "No Task\n",
      "Zero-Shot-CoT (left)\n",
      "Zero-...\n",
      "Metadata: {'page': 18, 'source': './data/2205.11916.pdf', 'total_pages': 42}\n",
      "Generated response length: 2383\n",
      "\n",
      "Processing query 13/15\n",
      "Query: How does the number of datasets and templates affect the performance of instruction tuning in the FL...\n",
      "Retrieved 4 documents\n",
      "First retrieved document:\n",
      "Length: 3112\n",
      "Content preview: Published as a conference paper at ICLR 2022\n",
      "B\n",
      "FURTHER ABLATION STUDIES AND ANALYSIS\n",
      "B.1\n",
      "DATASETS PER TASK CLUSTER & TEMPLATES PER DATASET\n",
      "Our primary hypothesis is that instruction tuning on a divers...\n",
      "Metadata: {'page': 24, 'source': './data/2109.01652.pdf', 'total_pages': 46}\n",
      "Generated response length: 1449\n",
      "\n",
      "Processing query 14/15\n",
      "Query: What technique surpasses zero-shot large language models on diverse reasoning tasks without hand-cra...\n",
      "Retrieved 4 documents\n",
      "First retrieved document:\n",
      "Length: 3166\n",
      "Content preview: Large Language Models are Zero-Shot Reasoners\n",
      "Takeshi Kojima\n",
      "The University of Tokyo\n",
      "t.kojima@weblab.t.u-tokyo.ac.jp\n",
      "Shixiang Shane Gu\n",
      "Google Research, Brain Team\n",
      "Machel Reid\n",
      "Google Research∗\n",
      "Yutaka M...\n",
      "Metadata: {'page': 1, 'source': './data/2205.11916.pdf', 'total_pages': 42}\n",
      "Generated response length: 1111\n",
      "\n",
      "Processing query 15/15\n",
      "Query: How does language model scale impact instruction tuning on zero-shot performance?...\n",
      "Retrieved 4 documents\n",
      "First retrieved document:\n",
      "Length: 3976\n",
      "Content preview: Published as a conference paper at ICLR 2022\n",
      "4.2\n",
      "SCALING LAWS\n",
      "0.4B\n",
      "2B\n",
      "8B\n",
      "68B\n",
      "137B\n",
      "Model Size (# parameters)\n",
      "30\n",
      "40\n",
      "50\n",
      "60\n",
      "70\n",
      "Instruction tuning\n",
      "Untuned model\n",
      "Average zero-shot accuracy \n",
      "on 13 held-out t...\n",
      "Metadata: {'page': 7, 'source': './data/2109.01652.pdf', 'total_pages': 46}\n",
      "Generated response length: 1397\n",
      "Successfully processed 15 queries\n",
      "\n",
      "Running evaluation metrics...\n"
     ]
    },
    {
     "data": {
      "application/vnd.jupyter.widget-view+json": {
       "model_id": "c9442eaf6f6f4512a027bbc92589eac8",
       "version_major": 2,
       "version_minor": 0
      },
      "text/plain": [
       "Evaluating:   0%|          | 0/60 [00:00<?, ?it/s]"
      ]
     },
     "metadata": {},
     "output_type": "display_data"
    },
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "trigger ascore single turn\n",
      "trigger ascore single turn\n",
      "trigger ascore single turn\n",
      "trigger ascore single turn\n",
      "trigger ascore single turn\n",
      "trigger ascore single turn\n",
      "trigger ascore single turn\n",
      "trigger ascore single turn\n",
      "trigger ascore single turn\n",
      "WRITE!\n",
      "WRITE!\n",
      "trigger ascore single turn\n",
      "WRITE!\n",
      "WRITE!\n",
      "trigger ascore single turn\n",
      "WRITE!\n",
      "trigger ascore single turn\n",
      "trigger ascore single turn\n",
      "WRITE!\n",
      "WRITE!\n",
      "trigger ascore single turn\n",
      "WRITE!\n",
      "trigger ascore single turn\n",
      "WRITE!\n",
      "WRITE!\n",
      "WRITE!\n",
      "WRITE!\n",
      "WRITE!\n",
      "WRITE!\n",
      "WRITE!\n",
      "Run 1 scores: {'context_recall': 1.0, 'factual_correctness': 0.11, 'faithfulness': 0.5384615384615384, 'semantic_similarity': 0.7629714645349647}\n",
      "\n",
      "Starting run 2/3\n",
      "Processing 15 valid queries\n",
      "\n",
      "Processing query 1/15\n",
      "Query: How does instruction tuning affect the zero-shot performance on unseen tasks?...\n",
      "Retrieved 4 documents\n",
      "First retrieved document:\n",
      "Length: 3976\n",
      "Content preview: Published as a conference paper at ICLR 2022\n",
      "4.2\n",
      "SCALING LAWS\n",
      "0.4B\n",
      "2B\n",
      "8B\n",
      "68B\n",
      "137B\n",
      "Model Size (# parameters)\n",
      "30\n",
      "40\n",
      "50\n",
      "60\n",
      "70\n",
      "Instruction tuning\n",
      "Untuned model\n",
      "Average zero-shot accuracy \n",
      "on 13 held-out t...\n",
      "Metadata: {'page': 7, 'source': './data/2109.01652.pdf', 'total_pages': 46}\n",
      "Generated response length: 3165\n",
      "\n",
      "Processing query 2/15\n",
      "Query: What is the Zero-shot-CoT method and how does it elicit chain of thought from large language models?...\n",
      "Retrieved 4 documents\n",
      "First retrieved document:\n",
      "Length: 3166\n",
      "Content preview: Large Language Models are Zero-Shot Reasoners\n",
      "Takeshi Kojima\n",
      "The University of Tokyo\n",
      "t.kojima@weblab.t.u-tokyo.ac.jp\n",
      "Shixiang Shane Gu\n",
      "Google Research, Brain Team\n",
      "Machel Reid\n",
      "Google Research∗\n",
      "Yutaka M...\n",
      "Metadata: {'page': 1, 'source': './data/2205.11916.pdf', 'total_pages': 42}\n",
      "Generated response length: 1884\n",
      "\n",
      "Processing query 3/15\n",
      "Query: How does prompt tuning affect model performance in improving NLP tasks?...\n",
      "Retrieved 4 documents\n",
      "First retrieved document:\n",
      "Length: 3925\n",
      "Content preview: Published as a conference paper at ICLR 2022\n",
      "instruct(xk)⊕yk ⊕instruct(x)”, where ⊕denotes string concatenation with a delimiter token inserted\n",
      "in between. At both training and inference time, exempla...\n",
      "Metadata: {'page': 8, 'source': './data/2109.01652.pdf', 'total_pages': 46}\n",
      "Generated response length: 2494\n",
      "\n",
      "Processing query 4/15\n",
      "Query: What is the purpose of instruction tuning in language models and how does it enhance response to NLP...\n",
      "Retrieved 4 documents\n",
      "First retrieved document:\n",
      "Length: 4454\n",
      "Content preview: Published as a conference paper at ICLR 2022\n",
      "et al., 2021). Compared with this prior work that ﬁnetunes and evaluates on the same downstream\n",
      "task, our setup studies the effect of instruction tuning on...\n",
      "Metadata: {'page': 29, 'source': './data/2109.01652.pdf', 'total_pages': 46}\n",
      "Generated response length: 2351\n",
      "\n",
      "Processing query 5/15\n",
      "Query: What distinguishes Zero-shot-CoT from Few-shot-CoT in terms of prompting methods for large language ...\n",
      "Retrieved 4 documents\n",
      "First retrieved document:\n",
      "Length: 3436\n",
      "Content preview: Table 4: Robustness study against template measured on the MultiArith dataset with text-davinci-002.\n",
      "(*1) This template is used in Ahn et al. [2022] where a language model is prompted to generate\n",
      "step...\n",
      "Metadata: {'page': 8, 'source': './data/2205.11916.pdf', 'total_pages': 42}\n",
      "Generated response length: 1937\n",
      "\n",
      "Processing query 6/15\n",
      "Query: Which language models were used in the experiment 'Exploring Zero-Shot Learning in Neural Networks',...\n",
      "Retrieved 4 documents\n",
      "First retrieved document:\n",
      "Length: 3052\n",
      "Content preview: Table 8: Description of language models. (*1) As for Original GPT3 models, we assign model size in-\n",
      "formation to each model by referring to https://blog.eleuther.ai/gpt3-model-sizes/ and\n",
      "https://beta....\n",
      "Metadata: {'page': 17, 'source': './data/2205.11916.pdf', 'total_pages': 42}\n",
      "Generated response length: 1465\n",
      "\n",
      "Processing query 7/15\n",
      "Query: How does Zero-shot-CoT differ from previous few-shot approaches in eliciting chain of thought from l...\n",
      "Retrieved 4 documents\n",
      "First retrieved document:\n",
      "Length: 4416\n",
      "Content preview: model (text-davinci-002). We also evaluate Zero-shot-CoT with another off-the-shelf large model,\n",
      "540B parameter PaLM, showing similar magnitudes of improvements on MultiArith and GSM8K.\n",
      "Importantly, w...\n",
      "Metadata: {'page': 3, 'source': './data/2205.11916.pdf', 'total_pages': 42}\n",
      "Generated response length: 2259\n",
      "\n",
      "Processing query 8/15\n",
      "Query: What are the stages in the Zero-shot-CoT method for reasoning and answer extraction, and how are the...\n",
      "Retrieved 4 documents\n",
      "First retrieved document:\n",
      "Length: 2726\n",
      "Content preview: Table 10: Answer extraction prompts used for Zero-shot-CoT experiments in Table 1. C.S.QA :\n",
      "CommonsenseQA, D.U. : Date Understanding, S.O. : Tracking Shufﬂed Objects\n",
      "No Task\n",
      "Zero-Shot-CoT (left)\n",
      "Zero-...\n",
      "Metadata: {'page': 18, 'source': './data/2205.11916.pdf', 'total_pages': 42}\n",
      "Generated response length: 2546\n",
      "\n",
      "Processing query 9/15\n",
      "Query: What are the main approaches for inducing LLMs to perform chain-of-thought reasoning, according to W...\n",
      "Retrieved 4 documents\n",
      "First retrieved document:\n",
      "Length: 4229\n",
      "Content preview: C\n",
      "Extended Related Work\n",
      "Chain-of-thought prompting is a general approach that is inspired by several prior directions: prompt-\n",
      "ing, natural language explanations, program synthesis/execution, numeric ...\n",
      "Metadata: {'page': 24, 'source': './data/2201.11903.pdf', 'total_pages': 43}\n",
      "Generated response length: 1880\n",
      "\n",
      "Processing query 10/15\n",
      "Query: Which sorting method has the most impact on Auto-CoT's demonstration accuracy in question clustering...\n",
      "Retrieved 4 documents\n",
      "First retrieved document:\n",
      "Length: 4087\n",
      "Content preview: the effect of misleading by similarity (Section 3.1). On the other hand, if we took each demonstration as a kind of skill,\n",
      "diverse demonstrations seem to cover more alternative skills for solving targ...\n",
      "Metadata: {'page': 6, 'source': './data/2210.03493.pdf', 'total_pages': 25}\n",
      "Generated response length: 2546\n",
      "\n",
      "Processing query 11/15\n",
      "Query: What are the pros and cons of prompting methods for large language models in terms of their cognitiv...\n",
      "Retrieved 4 documents\n",
      "First retrieved document:\n",
      "Length: 3166\n",
      "Content preview: Large Language Models are Zero-Shot Reasoners\n",
      "Takeshi Kojima\n",
      "The University of Tokyo\n",
      "t.kojima@weblab.t.u-tokyo.ac.jp\n",
      "Shixiang Shane Gu\n",
      "Google Research, Brain Team\n",
      "Machel Reid\n",
      "Google Research∗\n",
      "Yutaka M...\n",
      "Metadata: {'page': 1, 'source': './data/2205.11916.pdf', 'total_pages': 42}\n",
      "Generated response length: 3190\n",
      "\n",
      "Processing query 12/15\n",
      "Query: What are the stages in Zero-shot-CoT for reasoning and answer extraction, and how is it different fr...\n",
      "Retrieved 4 documents\n",
      "First retrieved document:\n",
      "Length: 2726\n",
      "Content preview: Table 10: Answer extraction prompts used for Zero-shot-CoT experiments in Table 1. C.S.QA :\n",
      "CommonsenseQA, D.U. : Date Understanding, S.O. : Tracking Shufﬂed Objects\n",
      "No Task\n",
      "Zero-Shot-CoT (left)\n",
      "Zero-...\n",
      "Metadata: {'page': 18, 'source': './data/2205.11916.pdf', 'total_pages': 42}\n",
      "Generated response length: 2002\n",
      "\n",
      "Processing query 13/15\n",
      "Query: How does the number of datasets and templates affect the performance of instruction tuning in the FL...\n",
      "Retrieved 4 documents\n",
      "First retrieved document:\n",
      "Length: 3112\n",
      "Content preview: Published as a conference paper at ICLR 2022\n",
      "B\n",
      "FURTHER ABLATION STUDIES AND ANALYSIS\n",
      "B.1\n",
      "DATASETS PER TASK CLUSTER & TEMPLATES PER DATASET\n",
      "Our primary hypothesis is that instruction tuning on a divers...\n",
      "Metadata: {'page': 24, 'source': './data/2109.01652.pdf', 'total_pages': 46}\n",
      "Generated response length: 2004\n",
      "\n",
      "Processing query 14/15\n",
      "Query: What technique surpasses zero-shot large language models on diverse reasoning tasks without hand-cra...\n",
      "Retrieved 4 documents\n",
      "First retrieved document:\n",
      "Length: 3166\n",
      "Content preview: Large Language Models are Zero-Shot Reasoners\n",
      "Takeshi Kojima\n",
      "The University of Tokyo\n",
      "t.kojima@weblab.t.u-tokyo.ac.jp\n",
      "Shixiang Shane Gu\n",
      "Google Research, Brain Team\n",
      "Machel Reid\n",
      "Google Research∗\n",
      "Yutaka M...\n",
      "Metadata: {'page': 1, 'source': './data/2205.11916.pdf', 'total_pages': 42}\n",
      "Generated response length: 1620\n",
      "\n",
      "Processing query 15/15\n",
      "Query: How does language model scale impact instruction tuning on zero-shot performance?...\n",
      "Retrieved 4 documents\n",
      "First retrieved document:\n",
      "Length: 3976\n",
      "Content preview: Published as a conference paper at ICLR 2022\n",
      "4.2\n",
      "SCALING LAWS\n",
      "0.4B\n",
      "2B\n",
      "8B\n",
      "68B\n",
      "137B\n",
      "Model Size (# parameters)\n",
      "30\n",
      "40\n",
      "50\n",
      "60\n",
      "70\n",
      "Instruction tuning\n",
      "Untuned model\n",
      "Average zero-shot accuracy \n",
      "on 13 held-out t...\n",
      "Metadata: {'page': 7, 'source': './data/2109.01652.pdf', 'total_pages': 46}\n",
      "Generated response length: 1266\n",
      "Successfully processed 15 queries\n",
      "\n",
      "Running evaluation metrics...\n"
     ]
    },
    {
     "data": {
      "application/vnd.jupyter.widget-view+json": {
       "model_id": "a323801332764ff6bce379007d55f7e6",
       "version_major": 2,
       "version_minor": 0
      },
      "text/plain": [
       "Evaluating:   0%|          | 0/60 [00:00<?, ?it/s]"
      ]
     },
     "metadata": {},
     "output_type": "display_data"
    },
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "trigger ascore single turn\n",
      "trigger ascore single turn\n",
      "trigger ascore single turn\n",
      "trigger ascore single turn\n",
      "trigger ascore single turn\n",
      "trigger ascore single turn\n",
      "trigger ascore single turn\n",
      "trigger ascore single turn\n",
      "trigger ascore single turn\n",
      "WRITE!\n",
      "WRITE!\n",
      "WRITE!\n",
      "trigger ascore single turn\n",
      "WRITE!\n",
      "trigger ascore single turn\n",
      "WRITE!\n",
      "trigger ascore single turn\n",
      "WRITE!\n",
      "WRITE!\n",
      "trigger ascore single turn\n",
      "trigger ascore single turn\n",
      "trigger ascore single turn\n",
      "WRITE!\n",
      "WRITE!\n",
      "WRITE!\n",
      "WRITE!\n",
      "WRITE!\n",
      "WRITE!\n",
      "WRITE!\n",
      "WRITE!\n",
      "Run 2 scores: {'context_recall': 1.0, 'factual_correctness': 0.0, 'faithfulness': 0.8333333333333334, 'semantic_similarity': 0.7736266932092302}\n",
      "\n",
      "Starting run 3/3\n",
      "Processing 15 valid queries\n",
      "\n",
      "Processing query 1/15\n",
      "Query: How does instruction tuning affect the zero-shot performance on unseen tasks?...\n",
      "Retrieved 4 documents\n",
      "First retrieved document:\n",
      "Length: 3976\n",
      "Content preview: Published as a conference paper at ICLR 2022\n",
      "4.2\n",
      "SCALING LAWS\n",
      "0.4B\n",
      "2B\n",
      "8B\n",
      "68B\n",
      "137B\n",
      "Model Size (# parameters)\n",
      "30\n",
      "40\n",
      "50\n",
      "60\n",
      "70\n",
      "Instruction tuning\n",
      "Untuned model\n",
      "Average zero-shot accuracy \n",
      "on 13 held-out t...\n",
      "Metadata: {'page': 7, 'source': './data/2109.01652.pdf', 'total_pages': 46}\n",
      "Generated response length: 3492\n",
      "\n",
      "Processing query 2/15\n",
      "Query: What is the Zero-shot-CoT method and how does it elicit chain of thought from large language models?...\n",
      "Retrieved 4 documents\n",
      "First retrieved document:\n",
      "Length: 3166\n",
      "Content preview: Large Language Models are Zero-Shot Reasoners\n",
      "Takeshi Kojima\n",
      "The University of Tokyo\n",
      "t.kojima@weblab.t.u-tokyo.ac.jp\n",
      "Shixiang Shane Gu\n",
      "Google Research, Brain Team\n",
      "Machel Reid\n",
      "Google Research∗\n",
      "Yutaka M...\n",
      "Metadata: {'page': 1, 'source': './data/2205.11916.pdf', 'total_pages': 42}\n",
      "Generated response length: 1779\n",
      "\n",
      "Processing query 3/15\n",
      "Query: How does prompt tuning affect model performance in improving NLP tasks?...\n",
      "Retrieved 4 documents\n",
      "First retrieved document:\n",
      "Length: 3925\n",
      "Content preview: Published as a conference paper at ICLR 2022\n",
      "instruct(xk)⊕yk ⊕instruct(x)”, where ⊕denotes string concatenation with a delimiter token inserted\n",
      "in between. At both training and inference time, exempla...\n",
      "Metadata: {'page': 8, 'source': './data/2109.01652.pdf', 'total_pages': 46}\n",
      "Generated response length: 2313\n",
      "\n",
      "Processing query 4/15\n",
      "Query: What is the purpose of instruction tuning in language models and how does it enhance response to NLP...\n",
      "Retrieved 4 documents\n",
      "First retrieved document:\n",
      "Length: 4454\n",
      "Content preview: Published as a conference paper at ICLR 2022\n",
      "et al., 2021). Compared with this prior work that ﬁnetunes and evaluates on the same downstream\n",
      "task, our setup studies the effect of instruction tuning on...\n",
      "Metadata: {'page': 29, 'source': './data/2109.01652.pdf', 'total_pages': 46}\n",
      "Generated response length: 2102\n",
      "\n",
      "Processing query 5/15\n",
      "Query: What distinguishes Zero-shot-CoT from Few-shot-CoT in terms of prompting methods for large language ...\n",
      "Retrieved 4 documents\n",
      "First retrieved document:\n",
      "Length: 3436\n",
      "Content preview: Table 4: Robustness study against template measured on the MultiArith dataset with text-davinci-002.\n",
      "(*1) This template is used in Ahn et al. [2022] where a language model is prompted to generate\n",
      "step...\n",
      "Metadata: {'page': 8, 'source': './data/2205.11916.pdf', 'total_pages': 42}\n",
      "Generated response length: 1766\n",
      "\n",
      "Processing query 6/15\n",
      "Query: Which language models were used in the experiment 'Exploring Zero-Shot Learning in Neural Networks',...\n",
      "Retrieved 4 documents\n",
      "First retrieved document:\n",
      "Length: 3052\n",
      "Content preview: Table 8: Description of language models. (*1) As for Original GPT3 models, we assign model size in-\n",
      "formation to each model by referring to https://blog.eleuther.ai/gpt3-model-sizes/ and\n",
      "https://beta....\n",
      "Metadata: {'page': 17, 'source': './data/2205.11916.pdf', 'total_pages': 42}\n",
      "Generated response length: 1605\n",
      "\n",
      "Processing query 7/15\n",
      "Query: How does Zero-shot-CoT differ from previous few-shot approaches in eliciting chain of thought from l...\n",
      "Retrieved 4 documents\n",
      "First retrieved document:\n",
      "Length: 4416\n",
      "Content preview: model (text-davinci-002). We also evaluate Zero-shot-CoT with another off-the-shelf large model,\n",
      "540B parameter PaLM, showing similar magnitudes of improvements on MultiArith and GSM8K.\n",
      "Importantly, w...\n",
      "Metadata: {'page': 3, 'source': './data/2205.11916.pdf', 'total_pages': 42}\n",
      "Generated response length: 1940\n",
      "\n",
      "Processing query 8/15\n",
      "Query: What are the stages in the Zero-shot-CoT method for reasoning and answer extraction, and how are the...\n",
      "Retrieved 4 documents\n",
      "First retrieved document:\n",
      "Length: 2726\n",
      "Content preview: Table 10: Answer extraction prompts used for Zero-shot-CoT experiments in Table 1. C.S.QA :\n",
      "CommonsenseQA, D.U. : Date Understanding, S.O. : Tracking Shufﬂed Objects\n",
      "No Task\n",
      "Zero-Shot-CoT (left)\n",
      "Zero-...\n",
      "Metadata: {'page': 18, 'source': './data/2205.11916.pdf', 'total_pages': 42}\n",
      "Generated response length: 1993\n",
      "\n",
      "Processing query 9/15\n",
      "Query: What are the main approaches for inducing LLMs to perform chain-of-thought reasoning, according to W...\n",
      "Retrieved 4 documents\n",
      "First retrieved document:\n",
      "Length: 4229\n",
      "Content preview: C\n",
      "Extended Related Work\n",
      "Chain-of-thought prompting is a general approach that is inspired by several prior directions: prompt-\n",
      "ing, natural language explanations, program synthesis/execution, numeric ...\n",
      "Metadata: {'page': 24, 'source': './data/2201.11903.pdf', 'total_pages': 43}\n",
      "Generated response length: 2014\n",
      "\n",
      "Processing query 10/15\n",
      "Query: Which sorting method has the most impact on Auto-CoT's demonstration accuracy in question clustering...\n",
      "Retrieved 4 documents\n",
      "First retrieved document:\n",
      "Length: 4087\n",
      "Content preview: the effect of misleading by similarity (Section 3.1). On the other hand, if we took each demonstration as a kind of skill,\n",
      "diverse demonstrations seem to cover more alternative skills for solving targ...\n",
      "Metadata: {'page': 6, 'source': './data/2210.03493.pdf', 'total_pages': 25}\n",
      "Generated response length: 2680\n",
      "\n",
      "Processing query 11/15\n",
      "Query: What are the pros and cons of prompting methods for large language models in terms of their cognitiv...\n",
      "Retrieved 4 documents\n",
      "First retrieved document:\n",
      "Length: 3166\n",
      "Content preview: Large Language Models are Zero-Shot Reasoners\n",
      "Takeshi Kojima\n",
      "The University of Tokyo\n",
      "t.kojima@weblab.t.u-tokyo.ac.jp\n",
      "Shixiang Shane Gu\n",
      "Google Research, Brain Team\n",
      "Machel Reid\n",
      "Google Research∗\n",
      "Yutaka M...\n",
      "Metadata: {'page': 1, 'source': './data/2205.11916.pdf', 'total_pages': 42}\n",
      "Generated response length: 2180\n",
      "\n",
      "Processing query 12/15\n",
      "Query: What are the stages in Zero-shot-CoT for reasoning and answer extraction, and how is it different fr...\n",
      "Retrieved 4 documents\n",
      "First retrieved document:\n",
      "Length: 2726\n",
      "Content preview: Table 10: Answer extraction prompts used for Zero-shot-CoT experiments in Table 1. C.S.QA :\n",
      "CommonsenseQA, D.U. : Date Understanding, S.O. : Tracking Shufﬂed Objects\n",
      "No Task\n",
      "Zero-Shot-CoT (left)\n",
      "Zero-...\n",
      "Metadata: {'page': 18, 'source': './data/2205.11916.pdf', 'total_pages': 42}\n",
      "Generated response length: 1989\n",
      "\n",
      "Processing query 13/15\n",
      "Query: How does the number of datasets and templates affect the performance of instruction tuning in the FL...\n",
      "Retrieved 4 documents\n",
      "First retrieved document:\n",
      "Length: 3112\n",
      "Content preview: Published as a conference paper at ICLR 2022\n",
      "B\n",
      "FURTHER ABLATION STUDIES AND ANALYSIS\n",
      "B.1\n",
      "DATASETS PER TASK CLUSTER & TEMPLATES PER DATASET\n",
      "Our primary hypothesis is that instruction tuning on a divers...\n",
      "Metadata: {'page': 24, 'source': './data/2109.01652.pdf', 'total_pages': 46}\n",
      "Generated response length: 1240\n",
      "\n",
      "Processing query 14/15\n",
      "Query: What technique surpasses zero-shot large language models on diverse reasoning tasks without hand-cra...\n",
      "Retrieved 4 documents\n",
      "First retrieved document:\n",
      "Length: 3166\n",
      "Content preview: Large Language Models are Zero-Shot Reasoners\n",
      "Takeshi Kojima\n",
      "The University of Tokyo\n",
      "t.kojima@weblab.t.u-tokyo.ac.jp\n",
      "Shixiang Shane Gu\n",
      "Google Research, Brain Team\n",
      "Machel Reid\n",
      "Google Research∗\n",
      "Yutaka M...\n",
      "Metadata: {'page': 1, 'source': './data/2205.11916.pdf', 'total_pages': 42}\n",
      "Generated response length: 1551\n",
      "\n",
      "Processing query 15/15\n",
      "Query: How does language model scale impact instruction tuning on zero-shot performance?...\n",
      "Retrieved 4 documents\n",
      "First retrieved document:\n",
      "Length: 3976\n",
      "Content preview: Published as a conference paper at ICLR 2022\n",
      "4.2\n",
      "SCALING LAWS\n",
      "0.4B\n",
      "2B\n",
      "8B\n",
      "68B\n",
      "137B\n",
      "Model Size (# parameters)\n",
      "30\n",
      "40\n",
      "50\n",
      "60\n",
      "70\n",
      "Instruction tuning\n",
      "Untuned model\n",
      "Average zero-shot accuracy \n",
      "on 13 held-out t...\n",
      "Metadata: {'page': 7, 'source': './data/2109.01652.pdf', 'total_pages': 46}\n",
      "Generated response length: 579\n",
      "Successfully processed 15 queries\n",
      "\n",
      "Running evaluation metrics...\n"
     ]
    },
    {
     "data": {
      "application/vnd.jupyter.widget-view+json": {
       "model_id": "7e803c13c4954039b5b5c91e3de420de",
       "version_major": 2,
       "version_minor": 0
      },
      "text/plain": [
       "Evaluating:   0%|          | 0/60 [00:00<?, ?it/s]"
      ]
     },
     "metadata": {},
     "output_type": "display_data"
    },
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "trigger ascore single turn\n",
      "trigger ascore single turn\n",
      "trigger ascore single turn\n",
      "trigger ascore single turn\n",
      "trigger ascore single turn\n",
      "trigger ascore single turn\n",
      "trigger ascore single turn\n",
      "trigger ascore single turn\n",
      "trigger ascore single turn\n",
      "WRITE!\n",
      "WRITE!\n",
      "trigger ascore single turn\n",
      "WRITE!\n",
      "WRITE!\n",
      "trigger ascore single turn\n",
      "WRITE!\n",
      "trigger ascore single turn\n",
      "trigger ascore single turn\n",
      "WRITE!\n",
      "trigger ascore single turn\n",
      "WRITE!\n",
      "trigger ascore single turn\n",
      "WRITE!\n",
      "WRITE!\n",
      "WRITE!\n",
      "WRITE!\n",
      "WRITE!\n",
      "WRITE!\n",
      "WRITE!\n",
      "WRITE!\n",
      "Run 3 scores: {'context_recall': 1.0, 'factual_correctness': 0.32, 'faithfulness': 0.7083333333333334, 'semantic_similarity': 0.7686114852060493}\n",
      "\n",
      "Final average scores for vector:\n",
      "context_recall: 1.0000 (95% CI: [1.0000, 1.0000])\n",
      "factual_correctness: 0.1433 (95% CI: [0.0106, 0.2761])\n",
      "faithfulness: 0.6934 (95% CI: [0.5725, 0.8142])\n",
      "semantic_similarity: 0.7684 (95% CI: [0.7641, 0.7728])\n",
      "\n",
      "Evaluating bm25...\n",
      "\n",
      "Running 3 evaluations for bm25...\n",
      "\n",
      "Starting run 1/3\n",
      "Processing 15 valid queries\n",
      "\n",
      "Processing query 1/15\n",
      "Query: How does instruction tuning affect the zero-shot performance on unseen tasks?...\n",
      "Retrieved 4 documents\n",
      "First retrieved document:\n",
      "Length: 4525\n",
      "Content preview: Published as a conference paper at ICLR 2022\n",
      "The way we ask a model to respond to instructions is similar to QA-based task formulation (Kumar\n",
      "et al., 2016; McCann et al., 2018), which aims to unify NL...\n",
      "Metadata: {'source': './data/2109.01652.pdf', 'page': 9, 'total_pages': 46}\n",
      "Generated response length: 1861\n",
      "\n",
      "Processing query 2/15\n",
      "Query: What is the Zero-shot-CoT method and how does it elicit chain of thought from large language models?...\n",
      "Retrieved 4 documents\n",
      "First retrieved document:\n",
      "Length: 4419\n",
      "Content preview: Sanh et al. [2022], Ouyang et al. [2022] show that such zero-shot abilities of LLMs can be increased\n",
      "by explicitly ﬁne-tuning models to follow instructions. Although these work focus on the zero-shot\n",
      "...\n",
      "Metadata: {'source': './data/2205.11916.pdf', 'page': 10, 'total_pages': 42}\n",
      "Generated response length: 2285\n",
      "\n",
      "Processing query 3/15\n",
      "Query: How does prompt tuning affect model performance in improving NLP tasks?...\n",
      "Retrieved 4 documents\n",
      "First retrieved document:\n",
      "Length: 4525\n",
      "Content preview: Published as a conference paper at ICLR 2022\n",
      "The way we ask a model to respond to instructions is similar to QA-based task formulation (Kumar\n",
      "et al., 2016; McCann et al., 2018), which aims to unify NL...\n",
      "Metadata: {'source': './data/2109.01652.pdf', 'page': 9, 'total_pages': 46}\n",
      "Generated response length: 2327\n",
      "\n",
      "Processing query 4/15\n",
      "Query: What is the purpose of instruction tuning in language models and how does it enhance response to NLP...\n",
      "Retrieved 4 documents\n",
      "First retrieved document:\n",
      "Length: 4525\n",
      "Content preview: Published as a conference paper at ICLR 2022\n",
      "The way we ask a model to respond to instructions is similar to QA-based task formulation (Kumar\n",
      "et al., 2016; McCann et al., 2018), which aims to unify NL...\n",
      "Metadata: {'source': './data/2109.01652.pdf', 'page': 9, 'total_pages': 46}\n",
      "Generated response length: 2404\n",
      "\n",
      "Processing query 5/15\n",
      "Query: What distinguishes Zero-shot-CoT from Few-shot-CoT in terms of prompting methods for large language ...\n",
      "Retrieved 4 documents\n",
      "First retrieved document:\n",
      "Length: 4416\n",
      "Content preview: model (text-davinci-002). We also evaluate Zero-shot-CoT with another off-the-shelf large model,\n",
      "540B parameter PaLM, showing similar magnitudes of improvements on MultiArith and GSM8K.\n",
      "Importantly, w...\n",
      "Metadata: {'source': './data/2205.11916.pdf', 'page': 3, 'total_pages': 42}\n",
      "Generated response length: 2072\n",
      "\n",
      "Processing query 6/15\n",
      "Query: Which language models were used in the experiment 'Exploring Zero-Shot Learning in Neural Networks',...\n",
      "Retrieved 4 documents\n",
      "First retrieved document:\n",
      "Length: 4430\n",
      "Content preview: Model Name\n",
      "nparams\n",
      "nlayers\n",
      "dmodel\n",
      "nheads\n",
      "dhead\n",
      "Batch Size\n",
      "Learning Rate\n",
      "GPT-3 Small\n",
      "125M\n",
      "12\n",
      "768\n",
      "12\n",
      "64\n",
      "0.5M\n",
      "6.0 × 10−4\n",
      "GPT-3 Medium\n",
      "350M\n",
      "24\n",
      "1024\n",
      "16\n",
      "64\n",
      "0.5M\n",
      "3.0 × 10−4\n",
      "GPT-3 Large\n",
      "760M\n",
      "24\n",
      "1536\n",
      "16\n",
      "96\n",
      "0.5...\n",
      "Metadata: {'source': './data/2005.14165.pdf', 'page': 8, 'total_pages': 75}\n",
      "Generated response length: 1465\n",
      "\n",
      "Processing query 7/15\n",
      "Query: How does Zero-shot-CoT differ from previous few-shot approaches in eliciting chain of thought from l...\n",
      "Retrieved 4 documents\n",
      "First retrieved document:\n",
      "Length: 4419\n",
      "Content preview: Sanh et al. [2022], Ouyang et al. [2022] show that such zero-shot abilities of LLMs can be increased\n",
      "by explicitly ﬁne-tuning models to follow instructions. Although these work focus on the zero-shot\n",
      "...\n",
      "Metadata: {'source': './data/2205.11916.pdf', 'page': 10, 'total_pages': 42}\n",
      "Generated response length: 2374\n",
      "\n",
      "Processing query 8/15\n",
      "Query: What are the stages in the Zero-shot-CoT method for reasoning and answer extraction, and how are the...\n",
      "Retrieved 4 documents\n",
      "First retrieved document:\n",
      "Length: 4871\n",
      "Content preview: based on prior knowledge [Talmor et al., 2019]. StrategyQA requires models to infer an implicit\n",
      "multi-hop reasoning to answer questions [Geva et al., 2021].\n",
      "For symbolic reasoning, we use Last Letter ...\n",
      "Metadata: {'source': './data/2205.11916.pdf', 'page': 5, 'total_pages': 42}\n",
      "Generated response length: 2406\n",
      "\n",
      "Processing query 9/15\n",
      "Query: What are the main approaches for inducing LLMs to perform chain-of-thought reasoning, according to W...\n",
      "Retrieved 4 documents\n",
      "First retrieved document:\n",
      "Length: 4168\n",
      "Content preview: A: Let’s think step by step. There are 64 puppies. 28 of\n",
      "them were sold. This leaves 36 puppies. Each cage has\n",
      "4 puppies, so we need 9 cages.\n",
      "Therefore, the answer (arabic numerals) is\n",
      "Test Question\n",
      "G...\n",
      "Metadata: {'source': './data/2210.03493.pdf', 'page': 2, 'total_pages': 25}\n",
      "Generated response length: 1575\n",
      "\n",
      "Processing query 10/15\n",
      "Query: Which sorting method has the most impact on Auto-CoT's demonstration accuracy in question clustering...\n",
      "Retrieved 4 documents\n",
      "First retrieved document:\n",
      "Length: 3337\n",
      "Content preview: reasoning in LLMs [Wei et al., 2022a, Kojima et al., 2022]. The model is accessed via the OpenAI API. 5 Greedy\n",
      "decoding is used to generate the output. We set max_tokens = 256 and temperature = 0. Fol...\n",
      "Metadata: {'source': './data/2210.03493.pdf', 'page': 15, 'total_pages': 25}\n",
      "Generated response length: 2083\n",
      "\n",
      "Processing query 11/15\n",
      "Query: What are the pros and cons of prompting methods for large language models in terms of their cognitiv...\n",
      "Retrieved 4 documents\n",
      "First retrieved document:\n",
      "Length: 4419\n",
      "Content preview: Sanh et al. [2022], Ouyang et al. [2022] show that such zero-shot abilities of LLMs can be increased\n",
      "by explicitly ﬁne-tuning models to follow instructions. Although these work focus on the zero-shot\n",
      "...\n",
      "Metadata: {'source': './data/2205.11916.pdf', 'page': 10, 'total_pages': 42}\n",
      "Generated response length: 2871\n",
      "\n",
      "Processing query 12/15\n",
      "Query: What are the stages in Zero-shot-CoT for reasoning and answer extraction, and how is it different fr...\n",
      "Retrieved 4 documents\n",
      "First retrieved document:\n",
      "Length: 3436\n",
      "Content preview: Table 4: Robustness study against template measured on the MultiArith dataset with text-davinci-002.\n",
      "(*1) This template is used in Ahn et al. [2022] where a language model is prompted to generate\n",
      "step...\n",
      "Metadata: {'source': './data/2205.11916.pdf', 'page': 8, 'total_pages': 42}\n",
      "Generated response length: 2107\n",
      "\n",
      "Processing query 13/15\n",
      "Query: How does the number of datasets and templates affect the performance of instruction tuning in the FL...\n",
      "Retrieved 4 documents\n",
      "First retrieved document:\n",
      "Length: 3938\n",
      "Content preview: Published as a conference paper at ICLR 2022\n",
      "English and Romanian–English from WMT’16 (Bojar et al., 2016). Compared with GPT-3, FLAN\n",
      "outperforms zero-shot GPT-3 for all six evaluations, though it und...\n",
      "Metadata: {'source': './data/2109.01652.pdf', 'page': 6, 'total_pages': 46}\n",
      "Generated response length: 1801\n",
      "\n",
      "Processing query 14/15\n",
      "Query: What technique surpasses zero-shot large language models on diverse reasoning tasks without hand-cra...\n",
      "Retrieved 4 documents\n",
      "First retrieved document:\n",
      "Length: 3166\n",
      "Content preview: Large Language Models are Zero-Shot Reasoners\n",
      "Takeshi Kojima\n",
      "The University of Tokyo\n",
      "t.kojima@weblab.t.u-tokyo.ac.jp\n",
      "Shixiang Shane Gu\n",
      "Google Research, Brain Team\n",
      "Machel Reid\n",
      "Google Research∗\n",
      "Yutaka M...\n",
      "Metadata: {'source': './data/2205.11916.pdf', 'page': 1, 'total_pages': 42}\n",
      "Generated response length: 1020\n",
      "\n",
      "Processing query 15/15\n",
      "Query: How does language model scale impact instruction tuning on zero-shot performance?...\n",
      "Retrieved 4 documents\n",
      "First retrieved document:\n",
      "Length: 4525\n",
      "Content preview: Published as a conference paper at ICLR 2022\n",
      "The way we ask a model to respond to instructions is similar to QA-based task formulation (Kumar\n",
      "et al., 2016; McCann et al., 2018), which aims to unify NL...\n",
      "Metadata: {'source': './data/2109.01652.pdf', 'page': 9, 'total_pages': 46}\n",
      "Generated response length: 2232\n",
      "Successfully processed 15 queries\n",
      "\n",
      "Running evaluation metrics...\n"
     ]
    },
    {
     "data": {
      "application/vnd.jupyter.widget-view+json": {
       "model_id": "2f768fc71e8c487f8c16bddbba322fea",
       "version_major": 2,
       "version_minor": 0
      },
      "text/plain": [
       "Evaluating:   0%|          | 0/60 [00:00<?, ?it/s]"
      ]
     },
     "metadata": {},
     "output_type": "display_data"
    },
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "trigger ascore single turn\n",
      "trigger ascore single turn\n",
      "trigger ascore single turn\n",
      "trigger ascore single turn\n",
      "trigger ascore single turn\n",
      "trigger ascore single turn\n",
      "trigger ascore single turn\n",
      "trigger ascore single turn\n",
      "WRITE!\n",
      "trigger ascore single turn\n",
      "WRITE!\n",
      "WRITE!\n",
      "WRITE!\n",
      "trigger ascore single turn\n",
      "trigger ascore single turn\n",
      "WRITE!\n",
      "trigger ascore single turn\n",
      "WRITE!\n",
      "trigger ascore single turn\n",
      "WRITE!\n",
      "trigger ascore single turn\n",
      "trigger ascore single turn\n",
      "WRITE!\n",
      "WRITE!\n",
      "WRITE!\n",
      "WRITE!\n",
      "WRITE!\n",
      "WRITE!\n",
      "WRITE!\n",
      "WRITE!\n",
      "Run 1 scores: {'context_recall': 1.0, 'factual_correctness': 0.47, 'faithfulness': 0.8666666666666667, 'semantic_similarity': 0.7326799177261398}\n",
      "\n",
      "Starting run 2/3\n",
      "Processing 15 valid queries\n",
      "\n",
      "Processing query 1/15\n",
      "Query: How does instruction tuning affect the zero-shot performance on unseen tasks?...\n",
      "Retrieved 4 documents\n",
      "First retrieved document:\n",
      "Length: 4525\n",
      "Content preview: Published as a conference paper at ICLR 2022\n",
      "The way we ask a model to respond to instructions is similar to QA-based task formulation (Kumar\n",
      "et al., 2016; McCann et al., 2018), which aims to unify NL...\n",
      "Metadata: {'source': './data/2109.01652.pdf', 'page': 9, 'total_pages': 46}\n",
      "Generated response length: 2058\n",
      "\n",
      "Processing query 2/15\n",
      "Query: What is the Zero-shot-CoT method and how does it elicit chain of thought from large language models?...\n",
      "Retrieved 4 documents\n",
      "First retrieved document:\n",
      "Length: 4419\n",
      "Content preview: Sanh et al. [2022], Ouyang et al. [2022] show that such zero-shot abilities of LLMs can be increased\n",
      "by explicitly ﬁne-tuning models to follow instructions. Although these work focus on the zero-shot\n",
      "...\n",
      "Metadata: {'source': './data/2205.11916.pdf', 'page': 10, 'total_pages': 42}\n",
      "Generated response length: 2431\n",
      "\n",
      "Processing query 3/15\n",
      "Query: How does prompt tuning affect model performance in improving NLP tasks?...\n",
      "Retrieved 4 documents\n",
      "First retrieved document:\n",
      "Length: 4525\n",
      "Content preview: Published as a conference paper at ICLR 2022\n",
      "The way we ask a model to respond to instructions is similar to QA-based task formulation (Kumar\n",
      "et al., 2016; McCann et al., 2018), which aims to unify NL...\n",
      "Metadata: {'source': './data/2109.01652.pdf', 'page': 9, 'total_pages': 46}\n",
      "Generated response length: 2469\n",
      "\n",
      "Processing query 4/15\n",
      "Query: What is the purpose of instruction tuning in language models and how does it enhance response to NLP...\n",
      "Retrieved 4 documents\n",
      "First retrieved document:\n",
      "Length: 4525\n",
      "Content preview: Published as a conference paper at ICLR 2022\n",
      "The way we ask a model to respond to instructions is similar to QA-based task formulation (Kumar\n",
      "et al., 2016; McCann et al., 2018), which aims to unify NL...\n",
      "Metadata: {'source': './data/2109.01652.pdf', 'page': 9, 'total_pages': 46}\n",
      "Generated response length: 2087\n",
      "\n",
      "Processing query 5/15\n",
      "Query: What distinguishes Zero-shot-CoT from Few-shot-CoT in terms of prompting methods for large language ...\n",
      "Retrieved 4 documents\n",
      "First retrieved document:\n",
      "Length: 4416\n",
      "Content preview: model (text-davinci-002). We also evaluate Zero-shot-CoT with another off-the-shelf large model,\n",
      "540B parameter PaLM, showing similar magnitudes of improvements on MultiArith and GSM8K.\n",
      "Importantly, w...\n",
      "Metadata: {'source': './data/2205.11916.pdf', 'page': 3, 'total_pages': 42}\n",
      "Generated response length: 1900\n",
      "\n",
      "Processing query 6/15\n",
      "Query: Which language models were used in the experiment 'Exploring Zero-Shot Learning in Neural Networks',...\n",
      "Retrieved 4 documents\n",
      "First retrieved document:\n",
      "Length: 4430\n",
      "Content preview: Model Name\n",
      "nparams\n",
      "nlayers\n",
      "dmodel\n",
      "nheads\n",
      "dhead\n",
      "Batch Size\n",
      "Learning Rate\n",
      "GPT-3 Small\n",
      "125M\n",
      "12\n",
      "768\n",
      "12\n",
      "64\n",
      "0.5M\n",
      "6.0 × 10−4\n",
      "GPT-3 Medium\n",
      "350M\n",
      "24\n",
      "1024\n",
      "16\n",
      "64\n",
      "0.5M\n",
      "3.0 × 10−4\n",
      "GPT-3 Large\n",
      "760M\n",
      "24\n",
      "1536\n",
      "16\n",
      "96\n",
      "0.5...\n",
      "Metadata: {'source': './data/2005.14165.pdf', 'page': 8, 'total_pages': 75}\n",
      "Generated response length: 983\n",
      "\n",
      "Processing query 7/15\n",
      "Query: How does Zero-shot-CoT differ from previous few-shot approaches in eliciting chain of thought from l...\n",
      "Retrieved 4 documents\n",
      "First retrieved document:\n",
      "Length: 4419\n",
      "Content preview: Sanh et al. [2022], Ouyang et al. [2022] show that such zero-shot abilities of LLMs can be increased\n",
      "by explicitly ﬁne-tuning models to follow instructions. Although these work focus on the zero-shot\n",
      "...\n",
      "Metadata: {'source': './data/2205.11916.pdf', 'page': 10, 'total_pages': 42}\n",
      "Generated response length: 2450\n",
      "\n",
      "Processing query 8/15\n",
      "Query: What are the stages in the Zero-shot-CoT method for reasoning and answer extraction, and how are the...\n",
      "Retrieved 4 documents\n",
      "First retrieved document:\n",
      "Length: 4871\n",
      "Content preview: based on prior knowledge [Talmor et al., 2019]. StrategyQA requires models to infer an implicit\n",
      "multi-hop reasoning to answer questions [Geva et al., 2021].\n",
      "For symbolic reasoning, we use Last Letter ...\n",
      "Metadata: {'source': './data/2205.11916.pdf', 'page': 5, 'total_pages': 42}\n",
      "Generated response length: 2031\n",
      "\n",
      "Processing query 9/15\n",
      "Query: What are the main approaches for inducing LLMs to perform chain-of-thought reasoning, according to W...\n",
      "Retrieved 4 documents\n",
      "First retrieved document:\n",
      "Length: 4168\n",
      "Content preview: A: Let’s think step by step. There are 64 puppies. 28 of\n",
      "them were sold. This leaves 36 puppies. Each cage has\n",
      "4 puppies, so we need 9 cages.\n",
      "Therefore, the answer (arabic numerals) is\n",
      "Test Question\n",
      "G...\n",
      "Metadata: {'source': './data/2210.03493.pdf', 'page': 2, 'total_pages': 25}\n",
      "Generated response length: 1390\n",
      "\n",
      "Processing query 10/15\n",
      "Query: Which sorting method has the most impact on Auto-CoT's demonstration accuracy in question clustering...\n",
      "Retrieved 4 documents\n",
      "First retrieved document:\n",
      "Length: 3337\n",
      "Content preview: reasoning in LLMs [Wei et al., 2022a, Kojima et al., 2022]. The model is accessed via the OpenAI API. 5 Greedy\n",
      "decoding is used to generate the output. We set max_tokens = 256 and temperature = 0. Fol...\n",
      "Metadata: {'source': './data/2210.03493.pdf', 'page': 15, 'total_pages': 25}\n",
      "Generated response length: 2128\n",
      "\n",
      "Processing query 11/15\n",
      "Query: What are the pros and cons of prompting methods for large language models in terms of their cognitiv...\n",
      "Retrieved 4 documents\n",
      "First retrieved document:\n",
      "Length: 4419\n",
      "Content preview: Sanh et al. [2022], Ouyang et al. [2022] show that such zero-shot abilities of LLMs can be increased\n",
      "by explicitly ﬁne-tuning models to follow instructions. Although these work focus on the zero-shot\n",
      "...\n",
      "Metadata: {'source': './data/2205.11916.pdf', 'page': 10, 'total_pages': 42}\n",
      "Generated response length: 2250\n",
      "\n",
      "Processing query 12/15\n",
      "Query: What are the stages in Zero-shot-CoT for reasoning and answer extraction, and how is it different fr...\n",
      "Retrieved 4 documents\n",
      "First retrieved document:\n",
      "Length: 3436\n",
      "Content preview: Table 4: Robustness study against template measured on the MultiArith dataset with text-davinci-002.\n",
      "(*1) This template is used in Ahn et al. [2022] where a language model is prompted to generate\n",
      "step...\n",
      "Metadata: {'source': './data/2205.11916.pdf', 'page': 8, 'total_pages': 42}\n",
      "Generated response length: 2128\n",
      "\n",
      "Processing query 13/15\n",
      "Query: How does the number of datasets and templates affect the performance of instruction tuning in the FL...\n",
      "Retrieved 4 documents\n",
      "First retrieved document:\n",
      "Length: 3938\n",
      "Content preview: Published as a conference paper at ICLR 2022\n",
      "English and Romanian–English from WMT’16 (Bojar et al., 2016). Compared with GPT-3, FLAN\n",
      "outperforms zero-shot GPT-3 for all six evaluations, though it und...\n",
      "Metadata: {'source': './data/2109.01652.pdf', 'page': 6, 'total_pages': 46}\n",
      "Generated response length: 1897\n",
      "\n",
      "Processing query 14/15\n",
      "Query: What technique surpasses zero-shot large language models on diverse reasoning tasks without hand-cra...\n",
      "Retrieved 4 documents\n",
      "First retrieved document:\n",
      "Length: 3166\n",
      "Content preview: Large Language Models are Zero-Shot Reasoners\n",
      "Takeshi Kojima\n",
      "The University of Tokyo\n",
      "t.kojima@weblab.t.u-tokyo.ac.jp\n",
      "Shixiang Shane Gu\n",
      "Google Research, Brain Team\n",
      "Machel Reid\n",
      "Google Research∗\n",
      "Yutaka M...\n",
      "Metadata: {'source': './data/2205.11916.pdf', 'page': 1, 'total_pages': 42}\n",
      "Generated response length: 467\n",
      "\n",
      "Processing query 15/15\n",
      "Query: How does language model scale impact instruction tuning on zero-shot performance?...\n",
      "Retrieved 4 documents\n",
      "First retrieved document:\n",
      "Length: 4525\n",
      "Content preview: Published as a conference paper at ICLR 2022\n",
      "The way we ask a model to respond to instructions is similar to QA-based task formulation (Kumar\n",
      "et al., 2016; McCann et al., 2018), which aims to unify NL...\n",
      "Metadata: {'source': './data/2109.01652.pdf', 'page': 9, 'total_pages': 46}\n",
      "Generated response length: 2018\n",
      "Successfully processed 15 queries\n",
      "\n",
      "Running evaluation metrics...\n"
     ]
    },
    {
     "data": {
      "application/vnd.jupyter.widget-view+json": {
       "model_id": "f165da2f1866412badc9659a6299ec4a",
       "version_major": 2,
       "version_minor": 0
      },
      "text/plain": [
       "Evaluating:   0%|          | 0/60 [00:00<?, ?it/s]"
      ]
     },
     "metadata": {},
     "output_type": "display_data"
    },
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "trigger ascore single turn\n",
      "trigger ascore single turn\n",
      "trigger ascore single turn\n",
      "trigger ascore single turn\n",
      "trigger ascore single turn\n",
      "trigger ascore single turn\n",
      "trigger ascore single turn\n",
      "trigger ascore single turn\n",
      "WRITE!\n",
      "trigger ascore single turn\n",
      "WRITE!\n",
      "WRITE!\n",
      "trigger ascore single turn\n",
      "WRITE!\n",
      "trigger ascore single turn\n",
      "WRITE!\n",
      "trigger ascore single turn\n",
      "trigger ascore single turn\n",
      "trigger ascore single turn\n",
      "WRITE!\n",
      "WRITE!\n",
      "trigger ascore single turn\n",
      "WRITE!\n",
      "WRITE!\n",
      "WRITE!\n",
      "WRITE!\n",
      "WRITE!\n",
      "WRITE!\n",
      "WRITE!\n",
      "WRITE!\n",
      "Run 2 scores: {'context_recall': 1.0, 'factual_correctness': 0.0, 'faithfulness': 0.6, 'semantic_similarity': 0.7661627025856231}\n",
      "\n",
      "Starting run 3/3\n",
      "Processing 15 valid queries\n",
      "\n",
      "Processing query 1/15\n",
      "Query: How does instruction tuning affect the zero-shot performance on unseen tasks?...\n",
      "Retrieved 4 documents\n",
      "First retrieved document:\n",
      "Length: 4525\n",
      "Content preview: Published as a conference paper at ICLR 2022\n",
      "The way we ask a model to respond to instructions is similar to QA-based task formulation (Kumar\n",
      "et al., 2016; McCann et al., 2018), which aims to unify NL...\n",
      "Metadata: {'source': './data/2109.01652.pdf', 'page': 9, 'total_pages': 46}\n",
      "Generated response length: 1910\n",
      "\n",
      "Processing query 2/15\n",
      "Query: What is the Zero-shot-CoT method and how does it elicit chain of thought from large language models?...\n",
      "Retrieved 4 documents\n",
      "First retrieved document:\n",
      "Length: 4419\n",
      "Content preview: Sanh et al. [2022], Ouyang et al. [2022] show that such zero-shot abilities of LLMs can be increased\n",
      "by explicitly ﬁne-tuning models to follow instructions. Although these work focus on the zero-shot\n",
      "...\n",
      "Metadata: {'source': './data/2205.11916.pdf', 'page': 10, 'total_pages': 42}\n",
      "Generated response length: 2174\n",
      "\n",
      "Processing query 3/15\n",
      "Query: How does prompt tuning affect model performance in improving NLP tasks?...\n",
      "Retrieved 4 documents\n",
      "First retrieved document:\n",
      "Length: 4525\n",
      "Content preview: Published as a conference paper at ICLR 2022\n",
      "The way we ask a model to respond to instructions is similar to QA-based task formulation (Kumar\n",
      "et al., 2016; McCann et al., 2018), which aims to unify NL...\n",
      "Metadata: {'source': './data/2109.01652.pdf', 'page': 9, 'total_pages': 46}\n",
      "Generated response length: 2136\n",
      "\n",
      "Processing query 4/15\n",
      "Query: What is the purpose of instruction tuning in language models and how does it enhance response to NLP...\n",
      "Retrieved 4 documents\n",
      "First retrieved document:\n",
      "Length: 4525\n",
      "Content preview: Published as a conference paper at ICLR 2022\n",
      "The way we ask a model to respond to instructions is similar to QA-based task formulation (Kumar\n",
      "et al., 2016; McCann et al., 2018), which aims to unify NL...\n",
      "Metadata: {'source': './data/2109.01652.pdf', 'page': 9, 'total_pages': 46}\n",
      "Generated response length: 2432\n",
      "\n",
      "Processing query 5/15\n",
      "Query: What distinguishes Zero-shot-CoT from Few-shot-CoT in terms of prompting methods for large language ...\n",
      "Retrieved 4 documents\n",
      "First retrieved document:\n",
      "Length: 4416\n",
      "Content preview: model (text-davinci-002). We also evaluate Zero-shot-CoT with another off-the-shelf large model,\n",
      "540B parameter PaLM, showing similar magnitudes of improvements on MultiArith and GSM8K.\n",
      "Importantly, w...\n",
      "Metadata: {'source': './data/2205.11916.pdf', 'page': 3, 'total_pages': 42}\n",
      "Generated response length: 2195\n",
      "\n",
      "Processing query 6/15\n",
      "Query: Which language models were used in the experiment 'Exploring Zero-Shot Learning in Neural Networks',...\n",
      "Retrieved 4 documents\n",
      "First retrieved document:\n",
      "Length: 4430\n",
      "Content preview: Model Name\n",
      "nparams\n",
      "nlayers\n",
      "dmodel\n",
      "nheads\n",
      "dhead\n",
      "Batch Size\n",
      "Learning Rate\n",
      "GPT-3 Small\n",
      "125M\n",
      "12\n",
      "768\n",
      "12\n",
      "64\n",
      "0.5M\n",
      "6.0 × 10−4\n",
      "GPT-3 Medium\n",
      "350M\n",
      "24\n",
      "1024\n",
      "16\n",
      "64\n",
      "0.5M\n",
      "3.0 × 10−4\n",
      "GPT-3 Large\n",
      "760M\n",
      "24\n",
      "1536\n",
      "16\n",
      "96\n",
      "0.5...\n",
      "Metadata: {'source': './data/2005.14165.pdf', 'page': 8, 'total_pages': 75}\n",
      "Generated response length: 441\n",
      "\n",
      "Processing query 7/15\n",
      "Query: How does Zero-shot-CoT differ from previous few-shot approaches in eliciting chain of thought from l...\n",
      "Retrieved 4 documents\n",
      "First retrieved document:\n",
      "Length: 4419\n",
      "Content preview: Sanh et al. [2022], Ouyang et al. [2022] show that such zero-shot abilities of LLMs can be increased\n",
      "by explicitly ﬁne-tuning models to follow instructions. Although these work focus on the zero-shot\n",
      "...\n",
      "Metadata: {'source': './data/2205.11916.pdf', 'page': 10, 'total_pages': 42}\n",
      "Generated response length: 2608\n",
      "\n",
      "Processing query 8/15\n",
      "Query: What are the stages in the Zero-shot-CoT method for reasoning and answer extraction, and how are the...\n",
      "Retrieved 4 documents\n",
      "First retrieved document:\n",
      "Length: 4871\n",
      "Content preview: based on prior knowledge [Talmor et al., 2019]. StrategyQA requires models to infer an implicit\n",
      "multi-hop reasoning to answer questions [Geva et al., 2021].\n",
      "For symbolic reasoning, we use Last Letter ...\n",
      "Metadata: {'source': './data/2205.11916.pdf', 'page': 5, 'total_pages': 42}\n",
      "Generated response length: 2028\n",
      "\n",
      "Processing query 9/15\n",
      "Query: What are the main approaches for inducing LLMs to perform chain-of-thought reasoning, according to W...\n",
      "Retrieved 4 documents\n",
      "First retrieved document:\n",
      "Length: 4168\n",
      "Content preview: A: Let’s think step by step. There are 64 puppies. 28 of\n",
      "them were sold. This leaves 36 puppies. Each cage has\n",
      "4 puppies, so we need 9 cages.\n",
      "Therefore, the answer (arabic numerals) is\n",
      "Test Question\n",
      "G...\n",
      "Metadata: {'source': './data/2210.03493.pdf', 'page': 2, 'total_pages': 25}\n",
      "Generated response length: 2761\n",
      "\n",
      "Processing query 10/15\n",
      "Query: Which sorting method has the most impact on Auto-CoT's demonstration accuracy in question clustering...\n",
      "Retrieved 4 documents\n",
      "First retrieved document:\n",
      "Length: 3337\n",
      "Content preview: reasoning in LLMs [Wei et al., 2022a, Kojima et al., 2022]. The model is accessed via the OpenAI API. 5 Greedy\n",
      "decoding is used to generate the output. We set max_tokens = 256 and temperature = 0. Fol...\n",
      "Metadata: {'source': './data/2210.03493.pdf', 'page': 15, 'total_pages': 25}\n",
      "Generated response length: 424\n",
      "\n",
      "Processing query 11/15\n",
      "Query: What are the pros and cons of prompting methods for large language models in terms of their cognitiv...\n",
      "Retrieved 4 documents\n",
      "First retrieved document:\n",
      "Length: 4419\n",
      "Content preview: Sanh et al. [2022], Ouyang et al. [2022] show that such zero-shot abilities of LLMs can be increased\n",
      "by explicitly ﬁne-tuning models to follow instructions. Although these work focus on the zero-shot\n",
      "...\n",
      "Metadata: {'source': './data/2205.11916.pdf', 'page': 10, 'total_pages': 42}\n",
      "Generated response length: 2709\n",
      "\n",
      "Processing query 12/15\n",
      "Query: What are the stages in Zero-shot-CoT for reasoning and answer extraction, and how is it different fr...\n",
      "Retrieved 4 documents\n",
      "First retrieved document:\n",
      "Length: 3436\n",
      "Content preview: Table 4: Robustness study against template measured on the MultiArith dataset with text-davinci-002.\n",
      "(*1) This template is used in Ahn et al. [2022] where a language model is prompted to generate\n",
      "step...\n",
      "Metadata: {'source': './data/2205.11916.pdf', 'page': 8, 'total_pages': 42}\n",
      "Generated response length: 1795\n",
      "\n",
      "Processing query 13/15\n",
      "Query: How does the number of datasets and templates affect the performance of instruction tuning in the FL...\n",
      "Retrieved 4 documents\n",
      "First retrieved document:\n",
      "Length: 3938\n",
      "Content preview: Published as a conference paper at ICLR 2022\n",
      "English and Romanian–English from WMT’16 (Bojar et al., 2016). Compared with GPT-3, FLAN\n",
      "outperforms zero-shot GPT-3 for all six evaluations, though it und...\n",
      "Metadata: {'source': './data/2109.01652.pdf', 'page': 6, 'total_pages': 46}\n",
      "Generated response length: 1716\n",
      "\n",
      "Processing query 14/15\n",
      "Query: What technique surpasses zero-shot large language models on diverse reasoning tasks without hand-cra...\n",
      "Retrieved 4 documents\n",
      "First retrieved document:\n",
      "Length: 3166\n",
      "Content preview: Large Language Models are Zero-Shot Reasoners\n",
      "Takeshi Kojima\n",
      "The University of Tokyo\n",
      "t.kojima@weblab.t.u-tokyo.ac.jp\n",
      "Shixiang Shane Gu\n",
      "Google Research, Brain Team\n",
      "Machel Reid\n",
      "Google Research∗\n",
      "Yutaka M...\n",
      "Metadata: {'source': './data/2205.11916.pdf', 'page': 1, 'total_pages': 42}\n",
      "Generated response length: 467\n",
      "\n",
      "Processing query 15/15\n",
      "Query: How does language model scale impact instruction tuning on zero-shot performance?...\n",
      "Retrieved 4 documents\n",
      "First retrieved document:\n",
      "Length: 4525\n",
      "Content preview: Published as a conference paper at ICLR 2022\n",
      "The way we ask a model to respond to instructions is similar to QA-based task formulation (Kumar\n",
      "et al., 2016; McCann et al., 2018), which aims to unify NL...\n",
      "Metadata: {'source': './data/2109.01652.pdf', 'page': 9, 'total_pages': 46}\n",
      "Generated response length: 956\n",
      "Successfully processed 15 queries\n",
      "\n",
      "Running evaluation metrics...\n"
     ]
    },
    {
     "data": {
      "application/vnd.jupyter.widget-view+json": {
       "model_id": "100494e3d5494f7191a1f024a6048a1d",
       "version_major": 2,
       "version_minor": 0
      },
      "text/plain": [
       "Evaluating:   0%|          | 0/60 [00:00<?, ?it/s]"
      ]
     },
     "metadata": {},
     "output_type": "display_data"
    },
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "trigger ascore single turn\n",
      "trigger ascore single turn\n",
      "trigger ascore single turn\n",
      "trigger ascore single turn\n",
      "trigger ascore single turn\n",
      "trigger ascore single turn\n",
      "trigger ascore single turn\n",
      "trigger ascore single turn\n",
      "WRITE!\n",
      "trigger ascore single turn\n",
      "trigger ascore single turn\n",
      "WRITE!\n",
      "WRITE!\n",
      "WRITE!\n",
      "trigger ascore single turn\n",
      "WRITE!\n",
      "trigger ascore single turn\n",
      "WRITE!\n",
      "trigger ascore single turn\n",
      "WRITE!\n",
      "WRITE!\n",
      "trigger ascore single turn\n",
      "trigger ascore single turn\n",
      "WRITE!\n",
      "WRITE!\n",
      "WRITE!\n",
      "WRITE!\n",
      "WRITE!\n",
      "WRITE!\n",
      "WRITE!\n",
      "Run 3 scores: {'context_recall': 1.0, 'factual_correctness': 0.12, 'faithfulness': 1.0, 'semantic_similarity': 0.7663608647002687}\n",
      "\n",
      "Final average scores for bm25:\n",
      "context_recall: 1.0000 (95% CI: [1.0000, 1.0000])\n",
      "factual_correctness: 0.1967 (95% CI: [0.0000, 0.3961])\n",
      "faithfulness: 0.8222 (95% CI: [0.6559, 0.9885])\n",
      "semantic_similarity: 0.7551 (95% CI: [0.7392, 0.7709])\n",
      "\n",
      "Evaluating hybrid...\n",
      "\n",
      "Running 3 evaluations for hybrid...\n",
      "\n",
      "Starting run 1/3\n",
      "Processing 15 valid queries\n",
      "\n",
      "Processing query 1/15\n",
      "Query: How does instruction tuning affect the zero-shot performance on unseen tasks?...\n",
      "Vector retriever returned 4 docs\n",
      "BM25 retriever returned 4 docs\n",
      "Hybrid retriever returning 4 docs\n",
      "Top document scores: [0.5, 0.15, 0.15]\n",
      "Vector retriever returned 4 docs\n",
      "BM25 retriever returned 4 docs\n",
      "Hybrid retriever returning 4 docs\n",
      "Top document scores: [0.5, 0.15, 0.15]\n",
      "Retrieved 4 documents\n",
      "First retrieved document:\n",
      "Length: 3976\n",
      "Content preview: Published as a conference paper at ICLR 2022\n",
      "4.2\n",
      "SCALING LAWS\n",
      "0.4B\n",
      "2B\n",
      "8B\n",
      "68B\n",
      "137B\n",
      "Model Size (# parameters)\n",
      "30\n",
      "40\n",
      "50\n",
      "60\n",
      "70\n",
      "Instruction tuning\n",
      "Untuned model\n",
      "Average zero-shot accuracy \n",
      "on 13 held-out t...\n",
      "Metadata: {'page': 7, 'source': './data/2109.01652.pdf', 'total_pages': 46, 'hybrid_score': 0.5}\n",
      "Generated response length: 1590\n",
      "\n",
      "Processing query 2/15\n",
      "Query: What is the Zero-shot-CoT method and how does it elicit chain of thought from large language models?...\n",
      "Vector retriever returned 4 docs\n",
      "BM25 retriever returned 4 docs\n",
      "Hybrid retriever returning 4 docs\n",
      "Top document scores: [0.15, 0.5, 0.15]\n",
      "Vector retriever returned 4 docs\n",
      "BM25 retriever returned 4 docs\n",
      "Hybrid retriever returning 4 docs\n",
      "Top document scores: [0.15, 0.5, 0.15]\n",
      "Retrieved 4 documents\n",
      "First retrieved document:\n",
      "Length: 4416\n",
      "Content preview: model (text-davinci-002). We also evaluate Zero-shot-CoT with another off-the-shelf large model,\n",
      "540B parameter PaLM, showing similar magnitudes of improvements on MultiArith and GSM8K.\n",
      "Importantly, w...\n",
      "Metadata: {'page': 3, 'source': './data/2205.11916.pdf', 'total_pages': 42, 'hybrid_score': 0.15}\n",
      "Generated response length: 2185\n",
      "\n",
      "Processing query 3/15\n",
      "Query: How does prompt tuning affect model performance in improving NLP tasks?...\n",
      "Vector retriever returned 4 docs\n",
      "BM25 retriever returned 4 docs\n",
      "Hybrid retriever returning 4 docs\n",
      "Top document scores: [0.5, 0.15, 0.5]\n",
      "Vector retriever returned 4 docs\n",
      "BM25 retriever returned 4 docs\n",
      "Hybrid retriever returning 4 docs\n",
      "Top document scores: [0.5, 0.15, 0.5]\n",
      "Retrieved 4 documents\n",
      "First retrieved document:\n",
      "Length: 3925\n",
      "Content preview: Published as a conference paper at ICLR 2022\n",
      "instruct(xk)⊕yk ⊕instruct(x)”, where ⊕denotes string concatenation with a delimiter token inserted\n",
      "in between. At both training and inference time, exempla...\n",
      "Metadata: {'page': 8, 'source': './data/2109.01652.pdf', 'total_pages': 46, 'hybrid_score': 0.5}\n",
      "Generated response length: 1661\n",
      "\n",
      "Processing query 4/15\n",
      "Query: What is the purpose of instruction tuning in language models and how does it enhance response to NLP...\n",
      "Vector retriever returned 4 docs\n",
      "BM25 retriever returned 4 docs\n",
      "Hybrid retriever returning 4 docs\n",
      "Top document scores: [0.5, 0.5, 0.5]\n",
      "Vector retriever returned 4 docs\n",
      "BM25 retriever returned 4 docs\n",
      "Hybrid retriever returning 4 docs\n",
      "Top document scores: [0.5, 0.5, 0.5]\n",
      "Retrieved 4 documents\n",
      "First retrieved document:\n",
      "Length: 4454\n",
      "Content preview: Published as a conference paper at ICLR 2022\n",
      "et al., 2021). Compared with this prior work that ﬁnetunes and evaluates on the same downstream\n",
      "task, our setup studies the effect of instruction tuning on...\n",
      "Metadata: {'page': 29, 'source': './data/2109.01652.pdf', 'total_pages': 46, 'hybrid_score': 0.5}\n",
      "Generated response length: 2186\n",
      "\n",
      "Processing query 5/15\n",
      "Query: What distinguishes Zero-shot-CoT from Few-shot-CoT in terms of prompting methods for large language ...\n",
      "Vector retriever returned 4 docs\n",
      "BM25 retriever returned 4 docs\n",
      "Hybrid retriever returning 4 docs\n",
      "Top document scores: [0.15, 0.15, 0.5]\n",
      "Vector retriever returned 4 docs\n",
      "BM25 retriever returned 4 docs\n",
      "Hybrid retriever returning 4 docs\n",
      "Top document scores: [0.15, 0.15, 0.5]\n",
      "Retrieved 4 documents\n",
      "First retrieved document:\n",
      "Length: 4416\n",
      "Content preview: model (text-davinci-002). We also evaluate Zero-shot-CoT with another off-the-shelf large model,\n",
      "540B parameter PaLM, showing similar magnitudes of improvements on MultiArith and GSM8K.\n",
      "Importantly, w...\n",
      "Metadata: {'page': 3, 'source': './data/2205.11916.pdf', 'total_pages': 42, 'hybrid_score': 0.15}\n",
      "Generated response length: 1899\n",
      "\n",
      "Processing query 6/15\n",
      "Query: Which language models were used in the experiment 'Exploring Zero-Shot Learning in Neural Networks',...\n",
      "Vector retriever returned 4 docs\n",
      "BM25 retriever returned 4 docs\n",
      "Hybrid retriever returning 4 docs\n",
      "Top document scores: [0.15, 0.15, 0.15]\n",
      "Vector retriever returned 4 docs\n",
      "BM25 retriever returned 4 docs\n",
      "Hybrid retriever returning 4 docs\n",
      "Top document scores: [0.15, 0.15, 0.15]\n",
      "Retrieved 4 documents\n",
      "First retrieved document:\n",
      "Length: 4430\n",
      "Content preview: Model Name\n",
      "nparams\n",
      "nlayers\n",
      "dmodel\n",
      "nheads\n",
      "dhead\n",
      "Batch Size\n",
      "Learning Rate\n",
      "GPT-3 Small\n",
      "125M\n",
      "12\n",
      "768\n",
      "12\n",
      "64\n",
      "0.5M\n",
      "6.0 × 10−4\n",
      "GPT-3 Medium\n",
      "350M\n",
      "24\n",
      "1024\n",
      "16\n",
      "64\n",
      "0.5M\n",
      "3.0 × 10−4\n",
      "GPT-3 Large\n",
      "760M\n",
      "24\n",
      "1536\n",
      "16\n",
      "96\n",
      "0.5...\n",
      "Metadata: {'source': './data/2005.14165.pdf', 'page': 8, 'total_pages': 75, 'hybrid_score': 0.15}\n",
      "Generated response length: 558\n",
      "\n",
      "Processing query 7/15\n",
      "Query: How does Zero-shot-CoT differ from previous few-shot approaches in eliciting chain of thought from l...\n",
      "Vector retriever returned 4 docs\n",
      "BM25 retriever returned 4 docs\n",
      "Hybrid retriever returning 4 docs\n",
      "Top document scores: [0.5, 0.15, 0.15]\n",
      "Vector retriever returned 4 docs\n",
      "BM25 retriever returned 4 docs\n",
      "Hybrid retriever returning 4 docs\n",
      "Top document scores: [0.5, 0.15, 0.15]\n",
      "Retrieved 4 documents\n",
      "First retrieved document:\n",
      "Length: 4416\n",
      "Content preview: model (text-davinci-002). We also evaluate Zero-shot-CoT with another off-the-shelf large model,\n",
      "540B parameter PaLM, showing similar magnitudes of improvements on MultiArith and GSM8K.\n",
      "Importantly, w...\n",
      "Metadata: {'page': 3, 'source': './data/2205.11916.pdf', 'total_pages': 42, 'hybrid_score': 0.5}\n",
      "Generated response length: 2377\n",
      "\n",
      "Processing query 8/15\n",
      "Query: What are the stages in the Zero-shot-CoT method for reasoning and answer extraction, and how are the...\n",
      "Vector retriever returned 4 docs\n",
      "BM25 retriever returned 4 docs\n",
      "Hybrid retriever returning 4 docs\n",
      "Top document scores: [0.15, 0.5, 0.15]\n",
      "Vector retriever returned 4 docs\n",
      "BM25 retriever returned 4 docs\n",
      "Hybrid retriever returning 4 docs\n",
      "Top document scores: [0.15, 0.5, 0.15]\n",
      "Retrieved 4 documents\n",
      "First retrieved document:\n",
      "Length: 3180\n",
      "Content preview: Table 1: Accuracy comparison of Zero-shot-CoT with Zero-shot on each tasks. The values on the left\n",
      "side of each task are the results of using answer extraction prompts depending on answer format as\n",
      "de...\n",
      "Metadata: {'page': 6, 'source': './data/2205.11916.pdf', 'total_pages': 42, 'hybrid_score': 0.15}\n",
      "Generated response length: 2057\n",
      "\n",
      "Processing query 9/15\n",
      "Query: What are the main approaches for inducing LLMs to perform chain-of-thought reasoning, according to W...\n",
      "Vector retriever returned 4 docs\n",
      "BM25 retriever returned 4 docs\n",
      "Hybrid retriever returning 4 docs\n",
      "Top document scores: [0.5, 0.15, 0.15]\n",
      "Vector retriever returned 4 docs\n",
      "BM25 retriever returned 4 docs\n",
      "Hybrid retriever returning 4 docs\n",
      "Top document scores: [0.5, 0.15, 0.15]\n",
      "Retrieved 4 documents\n",
      "First retrieved document:\n",
      "Length: 4229\n",
      "Content preview: C\n",
      "Extended Related Work\n",
      "Chain-of-thought prompting is a general approach that is inspired by several prior directions: prompt-\n",
      "ing, natural language explanations, program synthesis/execution, numeric ...\n",
      "Metadata: {'page': 24, 'source': './data/2201.11903.pdf', 'total_pages': 43, 'hybrid_score': 0.5}\n",
      "Generated response length: 1740\n",
      "\n",
      "Processing query 10/15\n",
      "Query: Which sorting method has the most impact on Auto-CoT's demonstration accuracy in question clustering...\n",
      "Vector retriever returned 4 docs\n",
      "BM25 retriever returned 4 docs\n",
      "Hybrid retriever returning 4 docs\n",
      "Top document scores: [0.15, 0.15, 0.15]\n",
      "Vector retriever returned 4 docs\n",
      "BM25 retriever returned 4 docs\n",
      "Hybrid retriever returning 4 docs\n",
      "Top document scores: [0.15, 0.15, 0.15]\n",
      "Retrieved 4 documents\n",
      "First retrieved document:\n",
      "Length: 3337\n",
      "Content preview: reasoning in LLMs [Wei et al., 2022a, Kojima et al., 2022]. The model is accessed via the OpenAI API. 5 Greedy\n",
      "decoding is used to generate the output. We set max_tokens = 256 and temperature = 0. Fol...\n",
      "Metadata: {'source': './data/2210.03493.pdf', 'page': 15, 'total_pages': 25, 'hybrid_score': 0.15}\n",
      "Generated response length: 1434\n",
      "\n",
      "Processing query 11/15\n",
      "Query: What are the pros and cons of prompting methods for large language models in terms of their cognitiv...\n",
      "Vector retriever returned 4 docs\n",
      "BM25 retriever returned 4 docs\n",
      "Hybrid retriever returning 4 docs\n",
      "Top document scores: [0.15, 0.15, 0.15]\n",
      "Vector retriever returned 4 docs\n",
      "BM25 retriever returned 4 docs\n",
      "Hybrid retriever returning 4 docs\n",
      "Top document scores: [0.15, 0.15, 0.15]\n",
      "Retrieved 4 documents\n",
      "First retrieved document:\n",
      "Length: 4419\n",
      "Content preview: Sanh et al. [2022], Ouyang et al. [2022] show that such zero-shot abilities of LLMs can be increased\n",
      "by explicitly ﬁne-tuning models to follow instructions. Although these work focus on the zero-shot\n",
      "...\n",
      "Metadata: {'source': './data/2205.11916.pdf', 'page': 10, 'total_pages': 42, 'hybrid_score': 0.15}\n",
      "Generated response length: 2926\n",
      "\n",
      "Processing query 12/15\n",
      "Query: What are the stages in Zero-shot-CoT for reasoning and answer extraction, and how is it different fr...\n",
      "Vector retriever returned 4 docs\n",
      "BM25 retriever returned 4 docs\n",
      "Hybrid retriever returning 4 docs\n",
      "Top document scores: [0.15, 0.15, 0.5]\n",
      "Vector retriever returned 4 docs\n",
      "BM25 retriever returned 4 docs\n",
      "Hybrid retriever returning 4 docs\n",
      "Top document scores: [0.15, 0.15, 0.5]\n",
      "Retrieved 4 documents\n",
      "First retrieved document:\n",
      "Length: 3436\n",
      "Content preview: Table 4: Robustness study against template measured on the MultiArith dataset with text-davinci-002.\n",
      "(*1) This template is used in Ahn et al. [2022] where a language model is prompted to generate\n",
      "step...\n",
      "Metadata: {'page': 8, 'source': './data/2205.11916.pdf', 'total_pages': 42, 'hybrid_score': 0.15}\n",
      "Generated response length: 2183\n",
      "\n",
      "Processing query 13/15\n",
      "Query: How does the number of datasets and templates affect the performance of instruction tuning in the FL...\n",
      "Vector retriever returned 4 docs\n",
      "BM25 retriever returned 4 docs\n",
      "Hybrid retriever returning 4 docs\n",
      "Top document scores: [0.5, 0.5, 0.15]\n",
      "Vector retriever returned 4 docs\n",
      "BM25 retriever returned 4 docs\n",
      "Hybrid retriever returning 4 docs\n",
      "Top document scores: [0.5, 0.5, 0.15]\n",
      "Retrieved 4 documents\n",
      "First retrieved document:\n",
      "Length: 3112\n",
      "Content preview: Published as a conference paper at ICLR 2022\n",
      "B\n",
      "FURTHER ABLATION STUDIES AND ANALYSIS\n",
      "B.1\n",
      "DATASETS PER TASK CLUSTER & TEMPLATES PER DATASET\n",
      "Our primary hypothesis is that instruction tuning on a divers...\n",
      "Metadata: {'page': 24, 'source': './data/2109.01652.pdf', 'total_pages': 46, 'hybrid_score': 0.5}\n",
      "Generated response length: 1893\n",
      "\n",
      "Processing query 14/15\n",
      "Query: What technique surpasses zero-shot large language models on diverse reasoning tasks without hand-cra...\n",
      "Vector retriever returned 4 docs\n",
      "BM25 retriever returned 4 docs\n",
      "Hybrid retriever returning 4 docs\n",
      "Top document scores: [0.5, 0.15, 0.15]\n",
      "Vector retriever returned 4 docs\n",
      "BM25 retriever returned 4 docs\n",
      "Hybrid retriever returning 4 docs\n",
      "Top document scores: [0.5, 0.15, 0.15]\n",
      "Retrieved 4 documents\n",
      "First retrieved document:\n",
      "Length: 3166\n",
      "Content preview: Large Language Models are Zero-Shot Reasoners\n",
      "Takeshi Kojima\n",
      "The University of Tokyo\n",
      "t.kojima@weblab.t.u-tokyo.ac.jp\n",
      "Shixiang Shane Gu\n",
      "Google Research, Brain Team\n",
      "Machel Reid\n",
      "Google Research∗\n",
      "Yutaka M...\n",
      "Metadata: {'page': 1, 'source': './data/2205.11916.pdf', 'total_pages': 42, 'hybrid_score': 0.5}\n",
      "Generated response length: 467\n",
      "\n",
      "Processing query 15/15\n",
      "Query: How does language model scale impact instruction tuning on zero-shot performance?...\n",
      "Vector retriever returned 4 docs\n",
      "BM25 retriever returned 4 docs\n",
      "Hybrid retriever returning 4 docs\n",
      "Top document scores: [0.5, 0.5, 0.15]\n",
      "Vector retriever returned 4 docs\n",
      "BM25 retriever returned 4 docs\n",
      "Hybrid retriever returning 4 docs\n",
      "Top document scores: [0.5, 0.5, 0.15]\n",
      "Retrieved 4 documents\n",
      "First retrieved document:\n",
      "Length: 3976\n",
      "Content preview: Published as a conference paper at ICLR 2022\n",
      "4.2\n",
      "SCALING LAWS\n",
      "0.4B\n",
      "2B\n",
      "8B\n",
      "68B\n",
      "137B\n",
      "Model Size (# parameters)\n",
      "30\n",
      "40\n",
      "50\n",
      "60\n",
      "70\n",
      "Instruction tuning\n",
      "Untuned model\n",
      "Average zero-shot accuracy \n",
      "on 13 held-out t...\n",
      "Metadata: {'page': 7, 'source': './data/2109.01652.pdf', 'total_pages': 46, 'hybrid_score': 0.5}\n",
      "Generated response length: 1712\n",
      "Successfully processed 15 queries\n",
      "\n",
      "Running evaluation metrics...\n"
     ]
    },
    {
     "data": {
      "application/vnd.jupyter.widget-view+json": {
       "model_id": "1bf63a4a5e4e4aaba435960d8231e63e",
       "version_major": 2,
       "version_minor": 0
      },
      "text/plain": [
       "Evaluating:   0%|          | 0/60 [00:00<?, ?it/s]"
      ]
     },
     "metadata": {},
     "output_type": "display_data"
    },
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "trigger ascore single turn\n",
      "trigger ascore single turn\n",
      "trigger ascore single turn\n",
      "trigger ascore single turn\n",
      "trigger ascore single turn\n",
      "trigger ascore single turn\n",
      "trigger ascore single turn\n",
      "trigger ascore single turn\n",
      "WRITE!\n",
      "trigger ascore single turn\n",
      "WRITE!\n",
      "trigger ascore single turn\n",
      "WRITE!\n",
      "trigger ascore single turn\n",
      "WRITE!\n",
      "WRITE!\n",
      "trigger ascore single turn\n",
      "trigger ascore single turn\n",
      "trigger ascore single turn\n",
      "WRITE!\n",
      "WRITE!\n",
      "trigger ascore single turn\n",
      "WRITE!\n",
      "WRITE!\n",
      "WRITE!\n",
      "WRITE!\n",
      "WRITE!\n",
      "WRITE!\n",
      "WRITE!\n",
      "WRITE!\n",
      "Run 1 scores: {'context_recall': 1.0, 'factual_correctness': 0.77, 'faithfulness': 1.0, 'semantic_similarity': 0.78927390408791}\n",
      "\n",
      "Starting run 2/3\n",
      "Processing 15 valid queries\n",
      "\n",
      "Processing query 1/15\n",
      "Query: How does instruction tuning affect the zero-shot performance on unseen tasks?...\n",
      "Vector retriever returned 4 docs\n",
      "BM25 retriever returned 4 docs\n",
      "Hybrid retriever returning 4 docs\n",
      "Top document scores: [0.5, 0.15, 0.15]\n",
      "Vector retriever returned 4 docs\n",
      "BM25 retriever returned 4 docs\n",
      "Hybrid retriever returning 4 docs\n",
      "Top document scores: [0.5, 0.15, 0.15]\n",
      "Retrieved 4 documents\n",
      "First retrieved document:\n",
      "Length: 3976\n",
      "Content preview: Published as a conference paper at ICLR 2022\n",
      "4.2\n",
      "SCALING LAWS\n",
      "0.4B\n",
      "2B\n",
      "8B\n",
      "68B\n",
      "137B\n",
      "Model Size (# parameters)\n",
      "30\n",
      "40\n",
      "50\n",
      "60\n",
      "70\n",
      "Instruction tuning\n",
      "Untuned model\n",
      "Average zero-shot accuracy \n",
      "on 13 held-out t...\n",
      "Metadata: {'page': 7, 'source': './data/2109.01652.pdf', 'total_pages': 46, 'hybrid_score': 0.5}\n",
      "Generated response length: 2835\n",
      "\n",
      "Processing query 2/15\n",
      "Query: What is the Zero-shot-CoT method and how does it elicit chain of thought from large language models?...\n",
      "Vector retriever returned 4 docs\n",
      "BM25 retriever returned 4 docs\n",
      "Hybrid retriever returning 4 docs\n",
      "Top document scores: [0.15, 0.5, 0.15]\n",
      "Vector retriever returned 4 docs\n",
      "BM25 retriever returned 4 docs\n",
      "Hybrid retriever returning 4 docs\n",
      "Top document scores: [0.15, 0.5, 0.15]\n",
      "Retrieved 4 documents\n",
      "First retrieved document:\n",
      "Length: 4416\n",
      "Content preview: model (text-davinci-002). We also evaluate Zero-shot-CoT with another off-the-shelf large model,\n",
      "540B parameter PaLM, showing similar magnitudes of improvements on MultiArith and GSM8K.\n",
      "Importantly, w...\n",
      "Metadata: {'page': 3, 'source': './data/2205.11916.pdf', 'total_pages': 42, 'hybrid_score': 0.15}\n",
      "Generated response length: 1818\n",
      "\n",
      "Processing query 3/15\n",
      "Query: How does prompt tuning affect model performance in improving NLP tasks?...\n",
      "Vector retriever returned 4 docs\n",
      "BM25 retriever returned 4 docs\n",
      "Hybrid retriever returning 4 docs\n",
      "Top document scores: [0.5, 0.15, 0.5]\n",
      "Vector retriever returned 4 docs\n",
      "BM25 retriever returned 4 docs\n",
      "Hybrid retriever returning 4 docs\n",
      "Top document scores: [0.5, 0.15, 0.5]\n",
      "Retrieved 4 documents\n",
      "First retrieved document:\n",
      "Length: 3925\n",
      "Content preview: Published as a conference paper at ICLR 2022\n",
      "instruct(xk)⊕yk ⊕instruct(x)”, where ⊕denotes string concatenation with a delimiter token inserted\n",
      "in between. At both training and inference time, exempla...\n",
      "Metadata: {'page': 8, 'source': './data/2109.01652.pdf', 'total_pages': 46, 'hybrid_score': 0.5}\n",
      "Generated response length: 2683\n",
      "\n",
      "Processing query 4/15\n",
      "Query: What is the purpose of instruction tuning in language models and how does it enhance response to NLP...\n",
      "Vector retriever returned 4 docs\n",
      "BM25 retriever returned 4 docs\n",
      "Hybrid retriever returning 4 docs\n",
      "Top document scores: [0.5, 0.5, 0.5]\n",
      "Vector retriever returned 4 docs\n",
      "BM25 retriever returned 4 docs\n",
      "Hybrid retriever returning 4 docs\n",
      "Top document scores: [0.5, 0.5, 0.5]\n",
      "Retrieved 4 documents\n",
      "First retrieved document:\n",
      "Length: 4454\n",
      "Content preview: Published as a conference paper at ICLR 2022\n",
      "et al., 2021). Compared with this prior work that ﬁnetunes and evaluates on the same downstream\n",
      "task, our setup studies the effect of instruction tuning on...\n",
      "Metadata: {'page': 29, 'source': './data/2109.01652.pdf', 'total_pages': 46, 'hybrid_score': 0.5}\n",
      "Generated response length: 2222\n",
      "\n",
      "Processing query 5/15\n",
      "Query: What distinguishes Zero-shot-CoT from Few-shot-CoT in terms of prompting methods for large language ...\n",
      "Vector retriever returned 4 docs\n",
      "BM25 retriever returned 4 docs\n",
      "Hybrid retriever returning 4 docs\n",
      "Top document scores: [0.15, 0.15, 0.5]\n",
      "Vector retriever returned 4 docs\n",
      "BM25 retriever returned 4 docs\n",
      "Hybrid retriever returning 4 docs\n",
      "Top document scores: [0.15, 0.15, 0.5]\n",
      "Retrieved 4 documents\n",
      "First retrieved document:\n",
      "Length: 4416\n",
      "Content preview: model (text-davinci-002). We also evaluate Zero-shot-CoT with another off-the-shelf large model,\n",
      "540B parameter PaLM, showing similar magnitudes of improvements on MultiArith and GSM8K.\n",
      "Importantly, w...\n",
      "Metadata: {'page': 3, 'source': './data/2205.11916.pdf', 'total_pages': 42, 'hybrid_score': 0.15}\n",
      "Generated response length: 1871\n",
      "\n",
      "Processing query 6/15\n",
      "Query: Which language models were used in the experiment 'Exploring Zero-Shot Learning in Neural Networks',...\n",
      "Vector retriever returned 4 docs\n",
      "BM25 retriever returned 4 docs\n",
      "Hybrid retriever returning 4 docs\n",
      "Top document scores: [0.15, 0.15, 0.15]\n",
      "Vector retriever returned 4 docs\n",
      "BM25 retriever returned 4 docs\n",
      "Hybrid retriever returning 4 docs\n",
      "Top document scores: [0.15, 0.15, 0.15]\n",
      "Retrieved 4 documents\n",
      "First retrieved document:\n",
      "Length: 4430\n",
      "Content preview: Model Name\n",
      "nparams\n",
      "nlayers\n",
      "dmodel\n",
      "nheads\n",
      "dhead\n",
      "Batch Size\n",
      "Learning Rate\n",
      "GPT-3 Small\n",
      "125M\n",
      "12\n",
      "768\n",
      "12\n",
      "64\n",
      "0.5M\n",
      "6.0 × 10−4\n",
      "GPT-3 Medium\n",
      "350M\n",
      "24\n",
      "1024\n",
      "16\n",
      "64\n",
      "0.5M\n",
      "3.0 × 10−4\n",
      "GPT-3 Large\n",
      "760M\n",
      "24\n",
      "1536\n",
      "16\n",
      "96\n",
      "0.5...\n",
      "Metadata: {'source': './data/2005.14165.pdf', 'page': 8, 'total_pages': 75, 'hybrid_score': 0.15}\n",
      "Generated response length: 410\n",
      "\n",
      "Processing query 7/15\n",
      "Query: How does Zero-shot-CoT differ from previous few-shot approaches in eliciting chain of thought from l...\n",
      "Vector retriever returned 4 docs\n",
      "BM25 retriever returned 4 docs\n",
      "Hybrid retriever returning 4 docs\n",
      "Top document scores: [0.5, 0.15, 0.15]\n",
      "Vector retriever returned 4 docs\n",
      "BM25 retriever returned 4 docs\n",
      "Hybrid retriever returning 4 docs\n",
      "Top document scores: [0.5, 0.15, 0.15]\n",
      "Retrieved 4 documents\n",
      "First retrieved document:\n",
      "Length: 4416\n",
      "Content preview: model (text-davinci-002). We also evaluate Zero-shot-CoT with another off-the-shelf large model,\n",
      "540B parameter PaLM, showing similar magnitudes of improvements on MultiArith and GSM8K.\n",
      "Importantly, w...\n",
      "Metadata: {'page': 3, 'source': './data/2205.11916.pdf', 'total_pages': 42, 'hybrid_score': 0.5}\n",
      "Generated response length: 2376\n",
      "\n",
      "Processing query 8/15\n",
      "Query: What are the stages in the Zero-shot-CoT method for reasoning and answer extraction, and how are the...\n",
      "Vector retriever returned 4 docs\n",
      "BM25 retriever returned 4 docs\n",
      "Hybrid retriever returning 4 docs\n",
      "Top document scores: [0.15, 0.5, 0.15]\n",
      "Vector retriever returned 4 docs\n",
      "BM25 retriever returned 4 docs\n",
      "Hybrid retriever returning 4 docs\n",
      "Top document scores: [0.15, 0.5, 0.15]\n",
      "Retrieved 4 documents\n",
      "First retrieved document:\n",
      "Length: 3180\n",
      "Content preview: Table 1: Accuracy comparison of Zero-shot-CoT with Zero-shot on each tasks. The values on the left\n",
      "side of each task are the results of using answer extraction prompts depending on answer format as\n",
      "de...\n",
      "Metadata: {'page': 6, 'source': './data/2205.11916.pdf', 'total_pages': 42, 'hybrid_score': 0.15}\n",
      "Generated response length: 2441\n",
      "\n",
      "Processing query 9/15\n",
      "Query: What are the main approaches for inducing LLMs to perform chain-of-thought reasoning, according to W...\n",
      "Vector retriever returned 4 docs\n",
      "BM25 retriever returned 4 docs\n",
      "Hybrid retriever returning 4 docs\n",
      "Top document scores: [0.5, 0.15, 0.15]\n",
      "Vector retriever returned 4 docs\n",
      "BM25 retriever returned 4 docs\n",
      "Hybrid retriever returning 4 docs\n",
      "Top document scores: [0.5, 0.15, 0.15]\n",
      "Retrieved 4 documents\n",
      "First retrieved document:\n",
      "Length: 4229\n",
      "Content preview: C\n",
      "Extended Related Work\n",
      "Chain-of-thought prompting is a general approach that is inspired by several prior directions: prompt-\n",
      "ing, natural language explanations, program synthesis/execution, numeric ...\n",
      "Metadata: {'page': 24, 'source': './data/2201.11903.pdf', 'total_pages': 43, 'hybrid_score': 0.5}\n",
      "Generated response length: 1578\n",
      "\n",
      "Processing query 10/15\n",
      "Query: Which sorting method has the most impact on Auto-CoT's demonstration accuracy in question clustering...\n",
      "Vector retriever returned 4 docs\n",
      "BM25 retriever returned 4 docs\n",
      "Hybrid retriever returning 4 docs\n",
      "Top document scores: [0.15, 0.15, 0.15]\n",
      "Vector retriever returned 4 docs\n",
      "BM25 retriever returned 4 docs\n",
      "Hybrid retriever returning 4 docs\n",
      "Top document scores: [0.15, 0.15, 0.15]\n",
      "Retrieved 4 documents\n",
      "First retrieved document:\n",
      "Length: 3337\n",
      "Content preview: reasoning in LLMs [Wei et al., 2022a, Kojima et al., 2022]. The model is accessed via the OpenAI API. 5 Greedy\n",
      "decoding is used to generate the output. We set max_tokens = 256 and temperature = 0. Fol...\n",
      "Metadata: {'source': './data/2210.03493.pdf', 'page': 15, 'total_pages': 25, 'hybrid_score': 0.15}\n",
      "Generated response length: 2283\n",
      "\n",
      "Processing query 11/15\n",
      "Query: What are the pros and cons of prompting methods for large language models in terms of their cognitiv...\n",
      "Vector retriever returned 4 docs\n",
      "BM25 retriever returned 4 docs\n",
      "Hybrid retriever returning 4 docs\n",
      "Top document scores: [0.15, 0.15, 0.15]\n",
      "Vector retriever returned 4 docs\n",
      "BM25 retriever returned 4 docs\n",
      "Hybrid retriever returning 4 docs\n",
      "Top document scores: [0.15, 0.15, 0.15]\n",
      "Retrieved 4 documents\n",
      "First retrieved document:\n",
      "Length: 4419\n",
      "Content preview: Sanh et al. [2022], Ouyang et al. [2022] show that such zero-shot abilities of LLMs can be increased\n",
      "by explicitly ﬁne-tuning models to follow instructions. Although these work focus on the zero-shot\n",
      "...\n",
      "Metadata: {'source': './data/2205.11916.pdf', 'page': 10, 'total_pages': 42, 'hybrid_score': 0.15}\n",
      "Generated response length: 3410\n",
      "\n",
      "Processing query 12/15\n",
      "Query: What are the stages in Zero-shot-CoT for reasoning and answer extraction, and how is it different fr...\n",
      "Vector retriever returned 4 docs\n",
      "BM25 retriever returned 4 docs\n",
      "Hybrid retriever returning 4 docs\n",
      "Top document scores: [0.15, 0.15, 0.5]\n",
      "Vector retriever returned 4 docs\n",
      "BM25 retriever returned 4 docs\n",
      "Hybrid retriever returning 4 docs\n",
      "Top document scores: [0.15, 0.15, 0.5]\n",
      "Retrieved 4 documents\n",
      "First retrieved document:\n",
      "Length: 3436\n",
      "Content preview: Table 4: Robustness study against template measured on the MultiArith dataset with text-davinci-002.\n",
      "(*1) This template is used in Ahn et al. [2022] where a language model is prompted to generate\n",
      "step...\n",
      "Metadata: {'page': 8, 'source': './data/2205.11916.pdf', 'total_pages': 42, 'hybrid_score': 0.15}\n",
      "Generated response length: 2269\n",
      "\n",
      "Processing query 13/15\n",
      "Query: How does the number of datasets and templates affect the performance of instruction tuning in the FL...\n",
      "Vector retriever returned 4 docs\n",
      "BM25 retriever returned 4 docs\n",
      "Hybrid retriever returning 4 docs\n",
      "Top document scores: [0.5, 0.5, 0.15]\n",
      "Vector retriever returned 4 docs\n",
      "BM25 retriever returned 4 docs\n",
      "Hybrid retriever returning 4 docs\n",
      "Top document scores: [0.5, 0.5, 0.15]\n",
      "Retrieved 4 documents\n",
      "First retrieved document:\n",
      "Length: 3112\n",
      "Content preview: Published as a conference paper at ICLR 2022\n",
      "B\n",
      "FURTHER ABLATION STUDIES AND ANALYSIS\n",
      "B.1\n",
      "DATASETS PER TASK CLUSTER & TEMPLATES PER DATASET\n",
      "Our primary hypothesis is that instruction tuning on a divers...\n",
      "Metadata: {'page': 24, 'source': './data/2109.01652.pdf', 'total_pages': 46, 'hybrid_score': 0.5}\n",
      "Generated response length: 2110\n",
      "\n",
      "Processing query 14/15\n",
      "Query: What technique surpasses zero-shot large language models on diverse reasoning tasks without hand-cra...\n",
      "Vector retriever returned 4 docs\n",
      "BM25 retriever returned 4 docs\n",
      "Hybrid retriever returning 4 docs\n",
      "Top document scores: [0.5, 0.15, 0.15]\n",
      "Vector retriever returned 4 docs\n",
      "BM25 retriever returned 4 docs\n",
      "Hybrid retriever returning 4 docs\n",
      "Top document scores: [0.5, 0.15, 0.15]\n",
      "Retrieved 4 documents\n",
      "First retrieved document:\n",
      "Length: 3166\n",
      "Content preview: Large Language Models are Zero-Shot Reasoners\n",
      "Takeshi Kojima\n",
      "The University of Tokyo\n",
      "t.kojima@weblab.t.u-tokyo.ac.jp\n",
      "Shixiang Shane Gu\n",
      "Google Research, Brain Team\n",
      "Machel Reid\n",
      "Google Research∗\n",
      "Yutaka M...\n",
      "Metadata: {'page': 1, 'source': './data/2205.11916.pdf', 'total_pages': 42, 'hybrid_score': 0.5}\n",
      "Generated response length: 1154\n",
      "\n",
      "Processing query 15/15\n",
      "Query: How does language model scale impact instruction tuning on zero-shot performance?...\n",
      "Vector retriever returned 4 docs\n",
      "BM25 retriever returned 4 docs\n",
      "Hybrid retriever returning 4 docs\n",
      "Top document scores: [0.5, 0.5, 0.15]\n",
      "Vector retriever returned 4 docs\n",
      "BM25 retriever returned 4 docs\n",
      "Hybrid retriever returning 4 docs\n",
      "Top document scores: [0.5, 0.5, 0.15]\n",
      "Retrieved 4 documents\n",
      "First retrieved document:\n",
      "Length: 3976\n",
      "Content preview: Published as a conference paper at ICLR 2022\n",
      "4.2\n",
      "SCALING LAWS\n",
      "0.4B\n",
      "2B\n",
      "8B\n",
      "68B\n",
      "137B\n",
      "Model Size (# parameters)\n",
      "30\n",
      "40\n",
      "50\n",
      "60\n",
      "70\n",
      "Instruction tuning\n",
      "Untuned model\n",
      "Average zero-shot accuracy \n",
      "on 13 held-out t...\n",
      "Metadata: {'page': 7, 'source': './data/2109.01652.pdf', 'total_pages': 46, 'hybrid_score': 0.5}\n",
      "Generated response length: 1766\n",
      "Successfully processed 15 queries\n",
      "\n",
      "Running evaluation metrics...\n"
     ]
    },
    {
     "data": {
      "application/vnd.jupyter.widget-view+json": {
       "model_id": "dd533465a71343149b6e3b9475238e00",
       "version_major": 2,
       "version_minor": 0
      },
      "text/plain": [
       "Evaluating:   0%|          | 0/60 [00:00<?, ?it/s]"
      ]
     },
     "metadata": {},
     "output_type": "display_data"
    },
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "trigger ascore single turn\n",
      "trigger ascore single turn\n",
      "trigger ascore single turn\n",
      "trigger ascore single turn\n",
      "trigger ascore single turn\n",
      "trigger ascore single turn\n",
      "trigger ascore single turn\n",
      "trigger ascore single turn\n",
      "trigger ascore single turn\n",
      "WRITE!\n",
      "WRITE!\n",
      "trigger ascore single turn\n",
      "WRITE!\n",
      "trigger ascore single turn\n",
      "WRITE!\n",
      "WRITE!\n",
      "WRITE!\n",
      "trigger ascore single turn\n",
      "trigger ascore single turn\n",
      "trigger ascore single turn\n",
      "WRITE!\n",
      "trigger ascore single turn\n",
      "WRITE!\n",
      "WRITE!\n",
      "WRITE!\n",
      "WRITE!\n",
      "WRITE!\n",
      "WRITE!\n",
      "WRITE!\n",
      "WRITE!\n",
      "Run 2 scores: {'context_recall': 1.0, 'factual_correctness': 0.33, 'faithfulness': 0.8947368421052632, 'semantic_similarity': 0.7525044829510555}\n",
      "\n",
      "Starting run 3/3\n",
      "Processing 15 valid queries\n",
      "\n",
      "Processing query 1/15\n",
      "Query: How does instruction tuning affect the zero-shot performance on unseen tasks?...\n",
      "Vector retriever returned 4 docs\n",
      "BM25 retriever returned 4 docs\n",
      "Hybrid retriever returning 4 docs\n",
      "Top document scores: [0.5, 0.15, 0.15]\n",
      "Vector retriever returned 4 docs\n",
      "BM25 retriever returned 4 docs\n",
      "Hybrid retriever returning 4 docs\n",
      "Top document scores: [0.5, 0.15, 0.15]\n",
      "Retrieved 4 documents\n",
      "First retrieved document:\n",
      "Length: 3976\n",
      "Content preview: Published as a conference paper at ICLR 2022\n",
      "4.2\n",
      "SCALING LAWS\n",
      "0.4B\n",
      "2B\n",
      "8B\n",
      "68B\n",
      "137B\n",
      "Model Size (# parameters)\n",
      "30\n",
      "40\n",
      "50\n",
      "60\n",
      "70\n",
      "Instruction tuning\n",
      "Untuned model\n",
      "Average zero-shot accuracy \n",
      "on 13 held-out t...\n",
      "Metadata: {'page': 7, 'source': './data/2109.01652.pdf', 'total_pages': 46, 'hybrid_score': 0.5}\n",
      "Generated response length: 2452\n",
      "\n",
      "Processing query 2/15\n",
      "Query: What is the Zero-shot-CoT method and how does it elicit chain of thought from large language models?...\n",
      "Vector retriever returned 4 docs\n",
      "BM25 retriever returned 4 docs\n",
      "Hybrid retriever returning 4 docs\n",
      "Top document scores: [0.15, 0.5, 0.15]\n",
      "Vector retriever returned 4 docs\n",
      "BM25 retriever returned 4 docs\n",
      "Hybrid retriever returning 4 docs\n",
      "Top document scores: [0.15, 0.5, 0.15]\n",
      "Retrieved 4 documents\n",
      "First retrieved document:\n",
      "Length: 4416\n",
      "Content preview: model (text-davinci-002). We also evaluate Zero-shot-CoT with another off-the-shelf large model,\n",
      "540B parameter PaLM, showing similar magnitudes of improvements on MultiArith and GSM8K.\n",
      "Importantly, w...\n",
      "Metadata: {'page': 3, 'source': './data/2205.11916.pdf', 'total_pages': 42, 'hybrid_score': 0.15}\n",
      "Generated response length: 2200\n",
      "\n",
      "Processing query 3/15\n",
      "Query: How does prompt tuning affect model performance in improving NLP tasks?...\n",
      "Vector retriever returned 4 docs\n",
      "BM25 retriever returned 4 docs\n",
      "Hybrid retriever returning 4 docs\n",
      "Top document scores: [0.5, 0.15, 0.5]\n",
      "Vector retriever returned 4 docs\n",
      "BM25 retriever returned 4 docs\n",
      "Hybrid retriever returning 4 docs\n",
      "Top document scores: [0.5, 0.15, 0.5]\n",
      "Retrieved 4 documents\n",
      "First retrieved document:\n",
      "Length: 3925\n",
      "Content preview: Published as a conference paper at ICLR 2022\n",
      "instruct(xk)⊕yk ⊕instruct(x)”, where ⊕denotes string concatenation with a delimiter token inserted\n",
      "in between. At both training and inference time, exempla...\n",
      "Metadata: {'page': 8, 'source': './data/2109.01652.pdf', 'total_pages': 46, 'hybrid_score': 0.5}\n",
      "Generated response length: 2136\n",
      "\n",
      "Processing query 4/15\n",
      "Query: What is the purpose of instruction tuning in language models and how does it enhance response to NLP...\n",
      "Vector retriever returned 4 docs\n",
      "BM25 retriever returned 4 docs\n",
      "Hybrid retriever returning 4 docs\n",
      "Top document scores: [0.5, 0.5, 0.5]\n",
      "Vector retriever returned 4 docs\n",
      "BM25 retriever returned 4 docs\n",
      "Hybrid retriever returning 4 docs\n",
      "Top document scores: [0.5, 0.5, 0.5]\n",
      "Retrieved 4 documents\n",
      "First retrieved document:\n",
      "Length: 4454\n",
      "Content preview: Published as a conference paper at ICLR 2022\n",
      "et al., 2021). Compared with this prior work that ﬁnetunes and evaluates on the same downstream\n",
      "task, our setup studies the effect of instruction tuning on...\n",
      "Metadata: {'page': 29, 'source': './data/2109.01652.pdf', 'total_pages': 46, 'hybrid_score': 0.5}\n",
      "Generated response length: 2260\n",
      "\n",
      "Processing query 5/15\n",
      "Query: What distinguishes Zero-shot-CoT from Few-shot-CoT in terms of prompting methods for large language ...\n",
      "Vector retriever returned 4 docs\n",
      "BM25 retriever returned 4 docs\n",
      "Hybrid retriever returning 4 docs\n",
      "Top document scores: [0.15, 0.15, 0.5]\n",
      "Vector retriever returned 4 docs\n",
      "BM25 retriever returned 4 docs\n",
      "Hybrid retriever returning 4 docs\n",
      "Top document scores: [0.15, 0.15, 0.5]\n",
      "Retrieved 4 documents\n",
      "First retrieved document:\n",
      "Length: 4416\n",
      "Content preview: model (text-davinci-002). We also evaluate Zero-shot-CoT with another off-the-shelf large model,\n",
      "540B parameter PaLM, showing similar magnitudes of improvements on MultiArith and GSM8K.\n",
      "Importantly, w...\n",
      "Metadata: {'page': 3, 'source': './data/2205.11916.pdf', 'total_pages': 42, 'hybrid_score': 0.15}\n",
      "Generated response length: 1647\n",
      "\n",
      "Processing query 6/15\n",
      "Query: Which language models were used in the experiment 'Exploring Zero-Shot Learning in Neural Networks',...\n",
      "Vector retriever returned 4 docs\n",
      "BM25 retriever returned 4 docs\n",
      "Hybrid retriever returning 4 docs\n",
      "Top document scores: [0.15, 0.15, 0.15]\n",
      "Vector retriever returned 4 docs\n",
      "BM25 retriever returned 4 docs\n",
      "Hybrid retriever returning 4 docs\n",
      "Top document scores: [0.15, 0.15, 0.15]\n",
      "Retrieved 4 documents\n",
      "First retrieved document:\n",
      "Length: 4430\n",
      "Content preview: Model Name\n",
      "nparams\n",
      "nlayers\n",
      "dmodel\n",
      "nheads\n",
      "dhead\n",
      "Batch Size\n",
      "Learning Rate\n",
      "GPT-3 Small\n",
      "125M\n",
      "12\n",
      "768\n",
      "12\n",
      "64\n",
      "0.5M\n",
      "6.0 × 10−4\n",
      "GPT-3 Medium\n",
      "350M\n",
      "24\n",
      "1024\n",
      "16\n",
      "64\n",
      "0.5M\n",
      "3.0 × 10−4\n",
      "GPT-3 Large\n",
      "760M\n",
      "24\n",
      "1536\n",
      "16\n",
      "96\n",
      "0.5...\n",
      "Metadata: {'source': './data/2005.14165.pdf', 'page': 8, 'total_pages': 75, 'hybrid_score': 0.15}\n",
      "Generated response length: 2254\n",
      "\n",
      "Processing query 7/15\n",
      "Query: How does Zero-shot-CoT differ from previous few-shot approaches in eliciting chain of thought from l...\n",
      "Vector retriever returned 4 docs\n",
      "BM25 retriever returned 4 docs\n",
      "Hybrid retriever returning 4 docs\n",
      "Top document scores: [0.5, 0.15, 0.15]\n",
      "Vector retriever returned 4 docs\n",
      "BM25 retriever returned 4 docs\n",
      "Hybrid retriever returning 4 docs\n",
      "Top document scores: [0.5, 0.15, 0.15]\n",
      "Retrieved 4 documents\n",
      "First retrieved document:\n",
      "Length: 4416\n",
      "Content preview: model (text-davinci-002). We also evaluate Zero-shot-CoT with another off-the-shelf large model,\n",
      "540B parameter PaLM, showing similar magnitudes of improvements on MultiArith and GSM8K.\n",
      "Importantly, w...\n",
      "Metadata: {'page': 3, 'source': './data/2205.11916.pdf', 'total_pages': 42, 'hybrid_score': 0.5}\n",
      "Generated response length: 1706\n",
      "\n",
      "Processing query 8/15\n",
      "Query: What are the stages in the Zero-shot-CoT method for reasoning and answer extraction, and how are the...\n",
      "Vector retriever returned 4 docs\n",
      "BM25 retriever returned 4 docs\n",
      "Hybrid retriever returning 4 docs\n",
      "Top document scores: [0.15, 0.5, 0.15]\n",
      "Vector retriever returned 4 docs\n",
      "BM25 retriever returned 4 docs\n",
      "Hybrid retriever returning 4 docs\n",
      "Top document scores: [0.15, 0.5, 0.15]\n",
      "Retrieved 4 documents\n",
      "First retrieved document:\n",
      "Length: 3180\n",
      "Content preview: Table 1: Accuracy comparison of Zero-shot-CoT with Zero-shot on each tasks. The values on the left\n",
      "side of each task are the results of using answer extraction prompts depending on answer format as\n",
      "de...\n",
      "Metadata: {'page': 6, 'source': './data/2205.11916.pdf', 'total_pages': 42, 'hybrid_score': 0.15}\n",
      "Generated response length: 1841\n",
      "\n",
      "Processing query 9/15\n",
      "Query: What are the main approaches for inducing LLMs to perform chain-of-thought reasoning, according to W...\n",
      "Vector retriever returned 4 docs\n",
      "BM25 retriever returned 4 docs\n",
      "Hybrid retriever returning 4 docs\n",
      "Top document scores: [0.5, 0.15, 0.15]\n",
      "Vector retriever returned 4 docs\n",
      "BM25 retriever returned 4 docs\n",
      "Hybrid retriever returning 4 docs\n",
      "Top document scores: [0.5, 0.15, 0.15]\n",
      "Retrieved 4 documents\n",
      "First retrieved document:\n",
      "Length: 4229\n",
      "Content preview: C\n",
      "Extended Related Work\n",
      "Chain-of-thought prompting is a general approach that is inspired by several prior directions: prompt-\n",
      "ing, natural language explanations, program synthesis/execution, numeric ...\n",
      "Metadata: {'page': 24, 'source': './data/2201.11903.pdf', 'total_pages': 43, 'hybrid_score': 0.5}\n",
      "Generated response length: 1492\n",
      "\n",
      "Processing query 10/15\n",
      "Query: Which sorting method has the most impact on Auto-CoT's demonstration accuracy in question clustering...\n",
      "Vector retriever returned 4 docs\n",
      "BM25 retriever returned 4 docs\n",
      "Hybrid retriever returning 4 docs\n",
      "Top document scores: [0.15, 0.15, 0.15]\n",
      "Vector retriever returned 4 docs\n",
      "BM25 retriever returned 4 docs\n",
      "Hybrid retriever returning 4 docs\n",
      "Top document scores: [0.15, 0.15, 0.15]\n",
      "Retrieved 4 documents\n",
      "First retrieved document:\n",
      "Length: 3337\n",
      "Content preview: reasoning in LLMs [Wei et al., 2022a, Kojima et al., 2022]. The model is accessed via the OpenAI API. 5 Greedy\n",
      "decoding is used to generate the output. We set max_tokens = 256 and temperature = 0. Fol...\n",
      "Metadata: {'source': './data/2210.03493.pdf', 'page': 15, 'total_pages': 25, 'hybrid_score': 0.15}\n",
      "Generated response length: 2419\n",
      "\n",
      "Processing query 11/15\n",
      "Query: What are the pros and cons of prompting methods for large language models in terms of their cognitiv...\n",
      "Vector retriever returned 4 docs\n",
      "BM25 retriever returned 4 docs\n",
      "Hybrid retriever returning 4 docs\n",
      "Top document scores: [0.15, 0.15, 0.15]\n",
      "Vector retriever returned 4 docs\n",
      "BM25 retriever returned 4 docs\n",
      "Hybrid retriever returning 4 docs\n",
      "Top document scores: [0.15, 0.15, 0.15]\n",
      "Retrieved 4 documents\n",
      "First retrieved document:\n",
      "Length: 4419\n",
      "Content preview: Sanh et al. [2022], Ouyang et al. [2022] show that such zero-shot abilities of LLMs can be increased\n",
      "by explicitly ﬁne-tuning models to follow instructions. Although these work focus on the zero-shot\n",
      "...\n",
      "Metadata: {'source': './data/2205.11916.pdf', 'page': 10, 'total_pages': 42, 'hybrid_score': 0.15}\n",
      "Generated response length: 2431\n",
      "\n",
      "Processing query 12/15\n",
      "Query: What are the stages in Zero-shot-CoT for reasoning and answer extraction, and how is it different fr...\n",
      "Vector retriever returned 4 docs\n",
      "BM25 retriever returned 4 docs\n",
      "Hybrid retriever returning 4 docs\n",
      "Top document scores: [0.15, 0.15, 0.5]\n",
      "Vector retriever returned 4 docs\n",
      "BM25 retriever returned 4 docs\n",
      "Hybrid retriever returning 4 docs\n",
      "Top document scores: [0.15, 0.15, 0.5]\n",
      "Retrieved 4 documents\n",
      "First retrieved document:\n",
      "Length: 3436\n",
      "Content preview: Table 4: Robustness study against template measured on the MultiArith dataset with text-davinci-002.\n",
      "(*1) This template is used in Ahn et al. [2022] where a language model is prompted to generate\n",
      "step...\n",
      "Metadata: {'page': 8, 'source': './data/2205.11916.pdf', 'total_pages': 42, 'hybrid_score': 0.15}\n",
      "Generated response length: 2371\n",
      "\n",
      "Processing query 13/15\n",
      "Query: How does the number of datasets and templates affect the performance of instruction tuning in the FL...\n",
      "Vector retriever returned 4 docs\n",
      "BM25 retriever returned 4 docs\n",
      "Hybrid retriever returning 4 docs\n",
      "Top document scores: [0.5, 0.5, 0.15]\n",
      "Vector retriever returned 4 docs\n",
      "BM25 retriever returned 4 docs\n",
      "Hybrid retriever returning 4 docs\n",
      "Top document scores: [0.5, 0.5, 0.15]\n",
      "Retrieved 4 documents\n",
      "First retrieved document:\n",
      "Length: 3112\n",
      "Content preview: Published as a conference paper at ICLR 2022\n",
      "B\n",
      "FURTHER ABLATION STUDIES AND ANALYSIS\n",
      "B.1\n",
      "DATASETS PER TASK CLUSTER & TEMPLATES PER DATASET\n",
      "Our primary hypothesis is that instruction tuning on a divers...\n",
      "Metadata: {'page': 24, 'source': './data/2109.01652.pdf', 'total_pages': 46, 'hybrid_score': 0.5}\n",
      "Generated response length: 2015\n",
      "\n",
      "Processing query 14/15\n",
      "Query: What technique surpasses zero-shot large language models on diverse reasoning tasks without hand-cra...\n",
      "Vector retriever returned 4 docs\n",
      "BM25 retriever returned 4 docs\n",
      "Hybrid retriever returning 4 docs\n",
      "Top document scores: [0.5, 0.15, 0.15]\n",
      "Vector retriever returned 4 docs\n",
      "BM25 retriever returned 4 docs\n",
      "Hybrid retriever returning 4 docs\n",
      "Top document scores: [0.5, 0.15, 0.15]\n",
      "Retrieved 4 documents\n",
      "First retrieved document:\n",
      "Length: 3166\n",
      "Content preview: Large Language Models are Zero-Shot Reasoners\n",
      "Takeshi Kojima\n",
      "The University of Tokyo\n",
      "t.kojima@weblab.t.u-tokyo.ac.jp\n",
      "Shixiang Shane Gu\n",
      "Google Research, Brain Team\n",
      "Machel Reid\n",
      "Google Research∗\n",
      "Yutaka M...\n",
      "Metadata: {'page': 1, 'source': './data/2205.11916.pdf', 'total_pages': 42, 'hybrid_score': 0.5}\n",
      "Generated response length: 467\n",
      "\n",
      "Processing query 15/15\n",
      "Query: How does language model scale impact instruction tuning on zero-shot performance?...\n",
      "Vector retriever returned 4 docs\n",
      "BM25 retriever returned 4 docs\n",
      "Hybrid retriever returning 4 docs\n",
      "Top document scores: [0.5, 0.5, 0.15]\n",
      "Vector retriever returned 4 docs\n",
      "BM25 retriever returned 4 docs\n",
      "Hybrid retriever returning 4 docs\n",
      "Top document scores: [0.5, 0.5, 0.15]\n",
      "Retrieved 4 documents\n",
      "First retrieved document:\n",
      "Length: 3976\n",
      "Content preview: Published as a conference paper at ICLR 2022\n",
      "4.2\n",
      "SCALING LAWS\n",
      "0.4B\n",
      "2B\n",
      "8B\n",
      "68B\n",
      "137B\n",
      "Model Size (# parameters)\n",
      "30\n",
      "40\n",
      "50\n",
      "60\n",
      "70\n",
      "Instruction tuning\n",
      "Untuned model\n",
      "Average zero-shot accuracy \n",
      "on 13 held-out t...\n",
      "Metadata: {'page': 7, 'source': './data/2109.01652.pdf', 'total_pages': 46, 'hybrid_score': 0.5}\n",
      "Generated response length: 2222\n",
      "Successfully processed 15 queries\n",
      "\n",
      "Running evaluation metrics...\n"
     ]
    },
    {
     "data": {
      "application/vnd.jupyter.widget-view+json": {
       "model_id": "c90841a24fdf47fca5ee573c6fc962b7",
       "version_major": 2,
       "version_minor": 0
      },
      "text/plain": [
       "Evaluating:   0%|          | 0/60 [00:00<?, ?it/s]"
      ]
     },
     "metadata": {},
     "output_type": "display_data"
    },
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "trigger ascore single turn\n",
      "trigger ascore single turn\n",
      "trigger ascore single turn\n",
      "trigger ascore single turn\n",
      "trigger ascore single turn\n",
      "trigger ascore single turn\n",
      "trigger ascore single turn\n",
      "trigger ascore single turn\n",
      "WRITE!\n",
      "WRITE!\n",
      "trigger ascore single turn\n",
      "WRITE!\n",
      "trigger ascore single turn\n",
      "WRITE!\n",
      "trigger ascore single turn\n",
      "WRITE!\n",
      "WRITE!\n",
      "trigger ascore single turn\n",
      "trigger ascore single turn\n",
      "trigger ascore single turn\n",
      "WRITE!\n",
      "trigger ascore single turn\n",
      "WRITE!\n",
      "WRITE!\n",
      "WRITE!\n",
      "WRITE!\n",
      "WRITE!\n",
      "WRITE!\n",
      "WRITE!\n",
      "WRITE!\n",
      "Run 3 scores: {'context_recall': 1.0, 'factual_correctness': 0.24, 'faithfulness': 0.9333333333333333, 'semantic_similarity': 0.769410649327454}\n",
      "\n",
      "Final average scores for hybrid:\n",
      "context_recall: 1.0000 (95% CI: [1.0000, 1.0000])\n",
      "factual_correctness: 0.4467 (95% CI: [0.2151, 0.6782])\n",
      "faithfulness: 0.9427 (95% CI: [0.8992, 0.9862])\n",
      "semantic_similarity: 0.7704 (95% CI: [0.7554, 0.7854])\n",
      "\n",
      "Evaluation Results:\n",
      "==================================================\n",
      "\n",
      "Results for vector:\n",
      "\n",
      "Mean scores across all runs:\n",
      "context_recall: 1.0000 (95% CI: [1.0000, 1.0000])\n",
      "factual_correctness: 0.1433 (95% CI: [0.0106, 0.2761])\n",
      "faithfulness: 0.6934 (95% CI: [0.5725, 0.8142])\n",
      "semantic_similarity: 0.7684 (95% CI: [0.7641, 0.7728])\n",
      "------------------------------\n",
      "\n",
      "Results for bm25:\n",
      "\n",
      "Mean scores across all runs:\n",
      "context_recall: 1.0000 (95% CI: [1.0000, 1.0000])\n",
      "factual_correctness: 0.1967 (95% CI: [0.0000, 0.3961])\n",
      "faithfulness: 0.8222 (95% CI: [0.6559, 0.9885])\n",
      "semantic_similarity: 0.7551 (95% CI: [0.7392, 0.7709])\n",
      "------------------------------\n",
      "\n",
      "Results for hybrid:\n",
      "\n",
      "Mean scores across all runs:\n",
      "context_recall: 1.0000 (95% CI: [1.0000, 1.0000])\n",
      "factual_correctness: 0.4467 (95% CI: [0.2151, 0.6782])\n",
      "faithfulness: 0.9427 (95% CI: [0.8992, 0.9862])\n",
      "semantic_similarity: 0.7704 (95% CI: [0.7554, 0.7854])\n",
      "------------------------------\n"
     ]
    },
    {
     "data": {
      "image/png": "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",
      "text/plain": [
       "<Figure size 1000x600 with 1 Axes>"
      ]
     },
     "metadata": {},
     "output_type": "display_data"
    },
    {
     "data": {
      "image/png": "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",
      "text/plain": [
       "<Figure size 1000x600 with 1 Axes>"
      ]
     },
     "metadata": {},
     "output_type": "display_data"
    },
    {
     "data": {
      "image/png": "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",
      "text/plain": [
       "<Figure size 1000x600 with 1 Axes>"
      ]
     },
     "metadata": {},
     "output_type": "display_data"
    },
    {
     "data": {
      "image/png": "iVBORw0KGgoAAAANSUhEUgAAA90AAAJOCAYAAACqS2TfAAAAOXRFWHRTb2Z0d2FyZQBNYXRwbG90bGliIHZlcnNpb24zLjkuMiwgaHR0cHM6Ly9tYXRwbG90bGliLm9yZy8hTgPZAAAACXBIWXMAAA9hAAAPYQGoP6dpAACIa0lEQVR4nOzdd3gU5fr/8c/uhhRKEkoSqqFIx9B7VVBEELECFoqIehRBIipYQPFoLMgBEcUGYqEIIkgRpNroEoh0hAAiNUASWgjZnd8f/LJf1iSYQJbZnbxf15Vz3Gdmd+97d8Kde8ozNsMwDAEAAAAAgHxnNzsAAAAAAACsiqYbAAAAAAAvoekGAAAAAMBLaLoBAAAAAPASmm4AAAAAALyEphsAAAAAAC+h6QYAAAAAwEtougEAAAAA8BKabgAAAAAAvISmGwD80CuvvCKbzeYxVrFiRfXp0+dfn/v555/LZrNp7969+RbP3r17ZbPZ9Pnnn+fbayL3VqxYIZvNphUrVpgdCq5Su3bt1K5du2v6nt7YfrL7d8YbufFvDwB/QNMNALlks9ly9XPpH65Hjx5VQECAHnzwwRxf99SpUwoJCdFdd911DbK4OlOmTNGYMWPMDiNbK1as0F133aXSpUsrMDBQkZGRuv322zVr1iyzQ0MepKena+zYsapfv75CQ0MVHh6u2rVr69FHH9X27dvNDi9fbN26Va+88kq+7vjKjsvl0hdffKGmTZuqRIkSKlasmKpVq6ZevXpp9erVXn1vMy1YsECvvPKK2WEAgFuA2QEAgL/48ssvPR5/8cUXWrx4cZbxmjVruv87MjJSN998s+bMmaOzZ8+qcOHCWV531qxZSktLu2xjnhs7duyQ3e7dfalTpkzR5s2b9fTTT3uMR0dH69y5cypUqJBX3z8nI0aM0MiRI1W1alU99thjio6O1vHjx7VgwQLdfffd+vrrr3X//febEtu10KZNG507d06BgYFmh3LV7r77bv3www/q2bOn+vfvrwsXLmj79u2aN2+eWrRooRo1apgd4lXbunWrXn31VbVr104VK1b0WPbjjz/m2/sMHDhQ48eP1x133KEHHnhAAQEB2rFjh3744QdVrlxZzZo1k+Sd7eehhx5Sjx49FBQUlG+vmZ3s/u1ZsGCBxo8fT+MNwGfQdANALv2zKV69erUWL178r83yAw88oIULF+r7779Xjx49siyfMmWKwsLC1Llz56uKz9t/3F6OzWZTcHCwKe89c+ZMjRw5Uvfcc4+mTJni8cf3s88+q0WLFunChQumxOZtaWlpCgwMlN1uN+3zz0/r1q3TvHnz9Prrr+uFF17wWPb+++8rOTnZnMCuofxqfI8cOaIPPvhA/fv318cff+yxbMyYMTp27Jj7sTe2H4fDIYfDka+veamMjAy5XC4FBgZaYtsHYG2cXg4AXnbnnXeqSJEimjJlSpZlR48e1dKlS3XPPfcoKChIv/zyi+69915dd911CgoKUoUKFTR48GCdO3fuX98nu2u6t2zZoptuukkhISEqX768/vvf/8rlcmV57pw5c9S5c2eVLVtWQUFBqlKlil577TU5nU73Ou3atdP8+fO1b98+96n0mUfpcrquctmyZWrdurWKFCmi8PBw3XHHHdq2bZvHOpnXp//555/q06ePwsPDFRYWpr59++rs2bP/mvfLL7+sEiVKaOLEidkeae/YsaO6dOnifnz06FH169dPUVFRCg4OVt26dTV58mSP52TmM2rUKI0fP16VK1dW4cKFdcstt+ivv/6SYRh67bXXVL58eYWEhOiOO+7QiRMnPF6jYsWK6tKli3788UfVq1dPwcHBqlWrVpbT3U+cOKEhQ4bohhtuUNGiRRUaGqpOnTpp06ZNHutlXnc7bdo0vfTSSypXrpwKFy6s1NTUbK/J3bVrl+6++26VLl1awcHBKl++vHr06KGUlBT3OhkZGXrttddUpUoVBQUFqWLFinrhhRd0/vz5bHP59ddf1aRJEwUHB6ty5cr64osv/vX7yYvdu3dLklq2bJllmcPhUMmSJT3G/v77bz388MOKiopSUFCQateurYkTJ3qsk/nZfPPNN3r11VdVrlw5FStWTPfcc49SUlJ0/vx5Pf3004qMjFTRokXVt2/fLPlPmjRJN910kyIjIxUUFKRatWrpww8/zBJjbj6nzz//XPfee68k6cYbb8xyWUp21z2npaXplVdeUbVq1RQcHKwyZcrorrvucn9e2UlMTJRhGNl+ljabTZGRkVk+o0u3n3bt2qlOnTpKSEhQ27ZtVbhwYV1//fWaOXOmJOmnn35S06ZNFRISourVq2vJkiUe75GbuSPS09M1fPhwNWzYUGFhYSpSpIhat26t5cuXe6x36e/jmDFj3Nvr1q1bs/zb06dPH40fP96dZ+aPYRiqWLGi7rjjjixxpKWlKSwsTI899liOsQLA1eBINwB4WZEiRXTHHXdo5syZOnHihEqUKOFeNn36dDmdTj3wwAOSpBkzZujs2bP6z3/+o5IlS2rt2rUaN26cDhw4oBkzZuTpfQ8fPqwbb7xRGRkZGjp0qIoUKaKPP/5YISEhWdb9/PPPVbRoUcXGxqpo0aJatmyZhg8frtTUVL3zzjuSpBdffFEpKSk6cOCA/ve//0mSihYtmuP7L1myRJ06dVLlypX1yiuv6Ny5cxo3bpxatmypDRs2ZDmt9r777lOlSpUUFxenDRs26NNPP1VkZKTeeuutHN9j165d2r59ux5++GEVK1bsXz+Tc+fOqV27dvrzzz81YMAAVapUSTNmzFCfPn2UnJysQYMGeaz/9ddfKz09XU899ZROnDiht99+W/fdd59uuukmrVixQs8//7z+/PNPjRs3TkOGDMnS8O3atUvdu3fX448/rt69e2vSpEm69957tXDhQt18882SpD179mj27Nm69957ValSJR05ckQfffSR2rZtq61bt6ps2bIer/naa68pMDBQQ4YM0fnz57M9Mpqenq6OHTvq/Pnzeuqpp1S6dGn9/fffmjdvnpKTkxUWFiZJeuSRRzR58mTdc889euaZZ7RmzRrFxcVp27Zt+u677zxe888//9Q999yjfv36qXfv3po4caL69Omjhg0bqnbt2v/62edGdHS0+3Nv2bKlAgJy/jPlyJEjatasmWw2mwYMGKCIiAj98MMP6tevn1JTU7NcAhEXF6eQkBANHTrU/Z0VKlRIdrtdJ0+e1CuvvKLVq1fr888/V6VKlTR8+HD3cz/88EPVrl1bXbt2VUBAgObOnasnnnhCLpdLTz75ZJ4+pzZt2mjgwIF677339MILL7gvR7n0spRLOZ1OdenSRUuXLlWPHj00aNAgnTp1SosXL9bmzZtVpUqVy36WM2bM0L333pvtpS3/5uTJk+rSpYt69Oihe++9Vx9++KF69Oihr7/+Wk8//bQef/xx3X///XrnnXd0zz336K+//srV72Gm1NRUffrpp+5LCU6dOqXPPvtMHTt21Nq1a1WvXj2P9SdNmqS0tDQ9+uijCgoKUokSJbLsRHzsscd08ODBLJf+2Gw2Pfjgg3r77bez/Ds8d+5cpaamXvUlPgCQIwMAcEWefPJJI7f/jM6fP9+QZHz00Uce482aNTPKlStnOJ1OwzAM4+zZs1meGxcXZ9hsNmPfvn3usREjRmR57+joaKN3797ux08//bQhyVizZo177OjRo0ZYWJghyUhMTHSPZ/e+jz32mFG4cGEjLS3NPda5c2cjOjo6y7qJiYmGJGPSpEnusXr16hmRkZHG8ePH3WObNm0y7Ha70atXryy5PPzwwx6veeeddxolS5bM8l6XmjNnjiHJ+N///nfZ9TKNGTPGkGR89dVX7rH09HSjefPmRtGiRY3U1FSPfCIiIozk5GT3usOGDTMkGXXr1jUuXLjgHu/Zs6cRGBjo8VlFR0cbkoxvv/3WPZaSkmKUKVPGqF+/vnssLS3N/f1nSkxMNIKCgoyRI0e6x5YvX25IMipXrpzl+8pctnz5csMwDCM+Pt6QZMyYMSPHz2Ljxo2GJOORRx7xGB8yZIghyVi2bFmWXH7++Wf32NGjR42goCDjmWeeyfE98srlchlt27Y1JBlRUVFGz549jfHjx3ts+5n69etnlClTxkhKSvIY79GjhxEWFub+jDI/mzp16hjp6enu9Xr27GnYbDajU6dOHs9v3rx5lm08u9+Pjh07GpUrV/YYy+3nNGPGDI/v61Jt27Y12rZt6348ceJEQ5IxevToLOu6XK4sY5fq1auXIckoXry4ceeddxqjRo0ytm3blmW9f24/mXFIMqZMmeIe2759uyHJsNvtxurVq93jixYtyvL7P2nSpCz/zvwzt4yMDOP8+fMesZw8edKIiory+Pcg8/cxNDTUOHr0qMf62f3bk9O/zTt27DAkGR9++KHHeNeuXY2KFSv+6+cJAFeK08sB4Bq45ZZbFBER4XGKeWJiolavXq2ePXu6J0C79Cj0mTNnlJSUpBYtWsgwDMXHx+fpPRcsWKBmzZqpSZMm7rGIiAj3UfVLXfq+p06dUlJSklq3bq2zZ89e0YzRhw4d0saNG9WnTx+PI0oxMTG6+eabtWDBgizPefzxxz0et27dWsePH1dqamqO75O5LLdH1xYsWKDSpUurZ8+e7rFChQpp4MCBOn36tH766SeP9e+99173UWFJatq0qaSL1/dfehS2adOmSk9P199//+3x/LJly+rOO+90Pw4NDVWvXr0UHx+vw4cPS7p4LX7m9+90OnX8+HEVLVpU1atX14YNG7Lk0Lt372zPVrhUZsyLFi3K8RT9zO8gNjbWY/yZZ56RJM2fP99jvFatWmrdurX7cUREhKpXr649e/ZcNpa8sNlsWrRokf773/+qePHimjp1qp588klFR0ere/fu7mu6DcPQt99+q9tvv12GYSgpKcn907FjR6WkpGT57Hr16uVx+UHTpk1lGIYefvhhj/WaNm2qv/76SxkZGe6xSz/vlJQUJSUlqW3bttqzZ4/H6fpS/n9O3377rUqVKqWnnnoqy7J/3jbwnyZNmqT3339flSpV0nfffachQ4aoZs2aat++fZZtNTtFixb1mIeievXqCg8PV82aNd2/C9L//V7kNUeHw+E+U8PlcunEiRPKyMhQo0aNst327777bkVEROTpPS5VrVo1NW3aVF9//bV77MSJE/rhhx/0wAMP/OvnCQBXiqYbAPLR6dOndfjwYfdP5mRFAQEB6t69u3755Rf3H7uZDfilTfD+/fvdjWrRokUVERGhtm3bSlKWP+7/zb59+1S1atUs49WrV88ytmXLFt15550KCwtTaGioIiIi3Kda5vV9M987p/eqWbOmkpKSdObMGY/x6667zuNx8eLFJV08xTUnoaGhki7uKMhtXFWrVs0yy3vmqb2ZcecUU2YzW6FChWzH/xnr9ddfn+UP+WrVqkmS+1pXl8ul//3vf6pataqCgoJUqlQpRUREKCEhIdvPvlKlSv+aZ6VKlRQbG6tPP/1UpUqVUseOHTV+/HiP19u3b5/sdruuv/56j+eWLl1a4eHh//pZSBe/o8t9P5J07Ngxj9+J06dPX3b9oKAgvfjii9q2bZsOHjyoqVOnqlmzZvrmm280YMAA92smJyfr448/VkREhMdP3759JV28dv9y8V/uu3S5XB6f1W+//aYOHTq45yaIiIhwT/T2z+/oSj+nnOzevVvVq1e/7Kn2ObHb7XryySf1+++/KykpSXPmzFGnTp20bNmybCd1/Kfy5ctn2X7DwsJyvf3nxuTJkxUTE6Pg4GCVLFlSERERmj9//hVv+/+mV69e+u2339zb94wZM3ThwgU99NBDV/3aAJATmm4AyEejRo1SmTJl3D+NGzd2L3vwwQflcrk0depUSdLUqVNVq1Yt93WLTqdTN998s+bPn6/nn39es2fP1uLFi90TBGU3AVp+SE5OVtu2bbVp0yaNHDlSc+fO1eLFi93XUnvrff8pp5mODcPI8TmZt4/6448/rmlMVxJrTt544w3FxsaqTZs2+uqrr7Ro0SItXrxYtWvXzvaz/7ej3JneffddJSQk6IUXXtC5c+c0cOBA1a5dWwcOHPBYL7dH964058aNG3v8TowaNSpX7ydJZcqUUY8ePfTzzz+ratWq+uabb9yzVksXf6cWL16c7c8/JxC70u9y9+7dat++vZKSkjR69GjNnz9fixcv1uDBgyVl/f3Iz20jP5UsWVJdu3bVggUL1LZtW/36669Zdqz8k7e3/6+++kp9+vRRlSpV9Nlnn2nhwoVavHixbrrppqva9i+nR48eKlSokPto91dffaVGjRplu4MQAPILE6kBQD7q1auXWrVq5X586R+JTZs2VZUqVTRlyhTdfPPN2rJli15//XX38j/++EM7d+7U5MmT1atXL/f44sWLryiW6Oho7dq1K8v4jh07PB6vWLFCx48f16xZs9SmTRv3eGJiYpbn5rZBy5zE6Z/vJUnbt29XqVKlVKRIkVy91uVUq1ZN1atX15w5czR27NjLTuyWGVdCQoJcLpfH0e7MU+gz484vf/75pwzD8Pjcdu7cKUnuieRmzpypG2+8UZ999pnHc5OTk1WqVKmrev8bbrhBN9xwg1566SWtXLlSLVu21IQJE/Tf//5X0dHRcrlc2rVrl8ckXkeOHFFycnK+fRZff/21x+z7lStXzvNrFCpUSDExMdq1a5eSkpIUERGhYsWKyel0qkOHDvkSZ07mzp2r8+fP6/vvv/c4iv3PGbbzIi+nMVepUkVr1qzRhQsXsp2d/0o0atRIP/30kw4dOpTv23xezJw5U5UrV9asWbM8PpMRI0Zc1ete7vMtUaKEOnfurK+//loPPPCAfvvtN40ZM+aq3g8A/g1HugEgH1WuXFkdOnRw//zzaNsDDzyg+Ph4jRgxQjabTffff797WebRo0uPFhmGobFjx15RLLfddptWr16ttWvXuseOHTvmcT1jTu+bnp6uDz74IMtrFilSJFenm5cpU0b16tXT5MmTPe6tvHnzZv3444+67bbb8ppOjl599VUdP35cjzzyiMd1uJl+/PFHzZs3T9LFz+Tw4cOaPn26e3lGRobGjRunokWLuk/lzy8HDx70mAU8NTVVX3zxherVq6fSpUtLuvj5//MI4YwZM3J1zW1OUlNTs3wWN9xwg+x2u/t2WJnfwT8bjtGjR0vSVd83PlPLli09ficu13Tv2rVL+/fvzzKenJysVatWqXjx4oqIiJDD4dDdd9+tb7/9Vps3b86y/qX3oL5a2f1+pKSkaNKkSVf8mpk7nHJz3/G7775bSUlJev/997Msu9yR5cOHD2vr1q1ZxtPT07V06dJsLy241rL7bNesWaNVq1Zd1ev+2+f70EMPaevWrXr22WflcDhydao9AFwNjnQDwDX04IMPauTIkZozZ45atmzpcdusGjVqqEqVKhoyZIj+/vtvhYaG6ttvv73ia0Gfe+45ffnll7r11ls1aNAg9y3DMo/2ZmrRooWKFy+u3r17a+DAgbLZbPryyy+z/YO+YcOGmj59umJjY9W4cWMVLVpUt99+e7bv/84776hTp05q3ry5+vXr575lWFhYmF555ZUryik73bt31x9//KHXX39d8fHx6tmzp6Kjo3X8+HEtXLhQS5cudV8//+ijj+qjjz5Snz599Pvvv6tixYqaOXOm+2hXXm53lBvVqlVTv379tG7dOkVFRWnixIk6cuSIR8PWpUsXjRw5Un379lWLFi30xx9/6Ouvv76iI8KZli1bpgEDBujee+9VtWrVlJGRoS+//NLdrEpS3bp11bt3b3388cfuSwzWrl2ryZMnq1u3brrxxhuvOv+82rRpk+6//3516tRJrVu3VokSJfT3339r8uTJOnjwoMaMGeNu1N58800tX75cTZs2Vf/+/VWrVi2dOHFCGzZs0JIlS7LcN/1K3XLLLQoMDNTtt9+uxx57TKdPn9Ynn3yiyMhIHTp06Ipes169enI4HHrrrbeUkpKioKAg933A/6lXr1764osvFBsbq7Vr16p169Y6c+aMlixZoieeeCLb+05L0oEDB9SkSRPddNNNat++vUqXLq2jR49q6tSp2rRpk55++umrPpPianXp0kWzZs3SnXfeqc6dOysxMVETJkxQrVq1/vXa/8tp2LChJGngwIHq2LFjlsa6c+fOKlmypGbMmKFOnTpl+7kDQH6i6QaAa6hq1apq3Lix1q1bl2UW8UKFCmnu3LkaOHCg4uLiFBwcrDvvvFMDBgxQ3bp18/xeZcqU0fLly/XUU0/pzTffVMmSJfX444+rbNmy6tevn3u9kiVLat68eXrmmWf00ksvqXjx4nrwwQfVvn17dezY0eM1n3jiCW3cuFGTJk3S//73P0VHR+fYdHfo0EELFy7UiBEjNHz4cBUqVEht27bVW2+9lS8TIl3qv//9r2666Sa99957+vDDD3XixAkVL15czZo105w5c9S1a1dJF0/3X7FihYYOHarJkycrNTVV1atX16RJk9SnT598jUm6+H2PGzdOzz77rHbs2KFKlSpp+vTpHp/rCy+8oDNnzmjKlCmaPn26GjRooPnz52vo0KFX/L5169ZVx44dNXfuXP39998qXLiw6tatqx9++EHNmjVzr/fpp5+qcuXK+vzzz/Xdd9+pdOnSGjZs2FWf3nul2rRpo9dee00//PCDRo8erWPHjqlYsWKqX7++3nrrLfcOA0mKiorS2rVrNXLkSM2aNUsffPCBSpYsqdq1a1/23u55Vb16dc2cOVMvvfSShgwZotKlS+s///mPIiIissx8nlulS5fWhAkTFBcXp379+snpdGr58uXZNn8Oh0MLFizQ66+/rilTpujbb79VyZIl1apVK91www2XjXvMmDFasGCBPvjgAx05ckTBwcGqU6eOPvnkE49/A8zSp08fHT58WB999JEWLVqkWrVq6auvvtKMGTO0YsWKK37du+66S0899ZSmTZumr776SoZheDTdgYGB6t69uz744AMmUANwTdgMs2f2AADAgipWrKg6deq4T20H4DsGDx6szz77TIcPH1bhwoXNDgeAxXFNNwAAAAqMtLQ0ffXVV7r77rtpuAFcE5xeDgAAAMs7evSolixZopkzZ+r48eMaNGiQ2SEBKCBougEAAGB5W7du1QMPPKDIyEi99957qlevntkhASggTD29/Oeff9btt9+usmXLymazafbs2f/6nBUrVqhBgwYKCgrS9ddfr88//9zrcQIAkFd79+7lem7Ah7Rr106GYejIkSMaMGCA2eEAKEBMbbrPnDmjunXravz48blaPzExUZ07d9aNN96ojRs36umnn9YjjzyiRYsWeTlSAAAAAADyzmdmL7fZbPruu+/UrVu3HNd5/vnnNX/+fG3evNk91qNHDyUnJ2vhwoXXIEoAAAAAAHLPr67pXrVqlTp06OAx1rFjRz399NM5Puf8+fM6f/68+7HL5dKJEydUsmRJ2Ww2b4UKAAAAALAwwzB06tQplS1bVnZ7zieR+1XTffjwYUVFRXmMRUVFKTU1VefOnVNISEiW58TFxenVV1+9ViECAAAAAAqQv/76S+XLl89xuV813Vdi2LBhio2NdT9OSUnRddddp8TERIWGhkqS7Ha77Ha7XC6XXC6Xe93McafTqUvPws9p3OFwyGazKSMjwyMGh8MhSXI6nbkaDwgIkGEYHuM2m00OhyNLjDmNkxM5kRM5kRM5kRM5kRM5kRM5kZP3cjp9+rQqVKigYsWK6XL8qukuXbq0jhw54jF25MgRhYaGZnuUW5KCgoIUFBSUZbxEiRLuptsXOZ1O7dy5U9WqVXNvXAAAXCvUIQCAWfylBmWeUv5vly2bOnt5XjVv3lxLly71GFu8eLGaN29uUkTeYxiGUlJSPPamAABwrVCHAABmsVoNMrXpPn36tDZu3KiNGzdKunhLsI0bN2r//v2SLp4a3qtXL/f6jz/+uPbs2aPnnntO27dv1wcffKBvvvlGgwcPNiN8AAAAAAAuy9Sme/369apfv77q168vSYqNjVX9+vU1fPhwSdKhQ4fcDbgkVapUSfPnz9fixYtVt25dvfvuu/r000/VsWNHU+IHAAAAAOByfOY+3ddKamqqwsLClJKS4tPXdLtcLiUlJalUqVKXnX4eAABvoA4BAMziLzUot72lX02kVpDY7XZFRkaaHQYAoICiDgEAzGK1GuS7uw0KOKfTqU2bNmWZDh8AgGuBOgQAMIvVahBNt48yDEPnzp2zzIx9AAD/Qh0CAJjFajWIphsAAAAAAC+h6QYAAAAAwEtoun2Uw+FQjRo15HA4zA4FAFAAUYcAAGaxWg1i9nIfZbPZFB4ebnYYAIACijoEADCL1WoQR7p9VEZGhtatW6eMjAyzQwEAFEDUIQCAWaxWg2i6fZhVpsgHAPgn6hAAwCxWqkE03QAAAAAAeAlNNwAAAAAAXkLT7aMcDodiYmIsM2MfAMC/UIcAAGaxWg2i6fZhgYGBZocAACjAqEMAALNYqQbRdPsop9Op9evXW2oCAQCA/6AOAQDMYrUaRNMNAAAAAICX0HQDAAAAAOAlNN0AAAAAAHiJzTAMw+wgrqXU1FSFhYUpJSVFoaGhZoeTI8Mw5HQ65XA4ZLPZzA4HAFDAUIcAAGbxlxqU296SI90+LD093ewQAAAFGHUIAGAWK9Ugmm4f5XQ6lZCQYJkZ+wAA/oU6BAAwi9VqEE03AAAAAABeQtMNAAAAAICX0HT7MIfDYXYIAIACjDoEADCLlWoQs5cDAAAAAJBHzF7u5wzDUHJysgrYPhEAgI+gDgEAzGK1GkTT7aOcTqe2b99umRn7AAD+hToEADCL1WoQTTcAAAAAAF5C0w0AAAAAgJfQdPsom82mkJAQ2Ww2s0MBABRA1CEAgFmsVoOYvRwAAAAAgDxi9nI/53K5dPToUblcLrNDAQAUQNQhAIBZrFaDaLp9lMvl0p49eyyzoQEA/At1CABgFqvVIJpuAAAAAAC8hKYbAAAAAAAvoen2UTabTWFhYZaZsQ8A4F+oQwAAs1itBjF7OQAAAAAAecTs5X7O5XLpwIEDlpk8AADgX6hDAACzWK0G0XT7KKttaAAA/0IdAgCYxWo1iKYbAAAAAAAvoekGAAAAAMBLaLp9lN1uV0REhOx2viIAwLVHHQIAmMVqNSjA7ACQPbvdripVqpgdBgCggKIOAQDMYrUaZI1dBxbkcrm0e/duy0weAADwL9QhAIBZrFaDaLp9lMvl0rFjxyyzoQEA/At1CABgFqvVIJpuAAAAAAC8hKYbAAAAAAAvoen2UXa7XeXLl7fMjH0AAP9CHQIAmMVqNYjZy31U5oYGAIAZqEMAALNYrQZZY9eBBTmdTm3btk1Op9PsUAAABRB1CABgFqvVIJpuH2UYhlJSUmQYhtmhAAAKIOoQAMAsVqtBNN0AAAAAAHgJTTcAAAAAAF5C0+2j7Ha7KleubJkZ+wAA/oU6BAAwi9VqELOX+yi73a7IyEizwwAAFFDUIQCAWaxWg6yx68CCnE6nNm3aZJkZ+wAA/oU6BAAwi9VqEE23jzIMQ+fOnbPMjH0AAP9CHQIAmMVqNYimGwAAAAAAL6HpBgAAAADAS2i6fZTD4VCNGjXkcDjMDgUAUABRhwAAZrFaDWL2ch9ls9kUHh5udhgAgAKKOgQAMIvVahBHun1URkaG1q1bp4yMDLNDAQAUQNQhAIBZrFaDaLp9mFWmyAcA+CfqEADALFaqQTTdAAAAAAB4CU03AAAAAABeQtPtoxwOh2JiYiwzYx8AwL9QhwAAZrFaDaLp9mGBgYFmhwAAKMCoQwAAs1ipBtF0+yin06n169dbagIBAID/oA4BAMxitRpE0w0AAAAAgJfQdAMAAAAA4CU03QAAAAAAeInNMAzD7CCupdTUVIWFhSklJUWhoaFmh5MjwzDkdDrlcDhks9nMDgcAUMBQhwAAZvGXGpTb3pIj3T4sPT3d7BAAAAUYdQgAYBYr1SCabh/ldDqVkJBgmRn7AAD+hToEADCL1WoQTTcAAAAAAF5C0w0AAAAAgJfQdPswh8NhdggAgAKMOgQAMIuVahCzlwMAAAAAkEfMXu7nDMNQcnKyCtg+EQCAj6AOAQDMYrUaRNPto5xOp7Zv326ZGfsAAP6FOgQAMIvVahBNNwAAAAAAXkLTDQAAAACAl9B0+yibzaaQkBDZbDazQwEAFEDUIQCAWaxWg5i9HAAAAACAPGL2cj/ncrl09OhRuVwus0MBABRA1CEAwJUaP368KlasqODgYDVt2lRr167Ncd127drJZrNl+bn55pvdNSi75TabTe+88477dU6cOKEHHnhAoaGhCg8PV79+/XT69Gmv55obpjfdeflCJGnMmDGqXr26QkJCVKFCBQ0ePFhpaWnXKNprx+Vyac+ePfyxAwAwBXUIAHAlpk+frtjYWI0YMUIbNmxQ3bp11bFjRx09ejTb9WfNmqVDhw65fzZv3iyHw6FmzZq5a9Clyw8dOqSJEyfKZrPp7rvvdr/OAw88oC1btmjx4sWaN2+efv75Zz366KPXJOd/Y2rTndcvZMqUKRo6dKhGjBihbdu26bPPPtP06dP1wgsvXOPIAQAAAAD/NHr0aPXv3199+/ZVrVq1NGHCBBUuXFgTJ07Mdv0SJUqodOnS7p/FixercOHCuummm9zrXLq8dOnSmjNnjm688UZVrlxZkrRt2zYtXLhQn376qZo2bapWrVpp3LhxmjZtmg4ePHhN8r4cU5vuvH4hK1euVMuWLXX//ferYsWKuuWWW9SzZ89/PToOAAAAAPCu9PR0/f777+rQoYN7zG63q0OHDlq1alWuXuOzzz7Tfffdp5CQkGyXHzlyRPPnz1e/fv3cY6tWrVJ4eLgaNWrkHuvQoYPsdrvWrFlzhdnkH9Oa7iv5Qlq0aKHff//d3WTv2bNHCxYs0G233XZNYr6WbDabwsLCLDNjHwDAv1CHAAB5lZSUJKfTqaioKI/xqKgoHT58+F+fv3btWm3evFn9+vXLsQZNnjxZxYoV01133eUeO3z4sCIjIz3WCwgIUIkSJXL1vt4WYNYbX+4L2b59e7bPuf/++5WUlKRWrVrJMAxlZGTo8ccfv+zp5efPn9f58+fdj1NTUyVJGRkZysjIkHSx2bfb7XK5XB7XrmWOO51OXTrJe07jDodDNpvN/bqXjkuS0+nM1XhAQIDsdruqVq3qztNms8nhcGSJMadxX8zJMAyPcXIiJ3IiJ3Ly3Zz+WYeskFN24+RETuRETuSUfzll/rfT6fSIMzOPf8vpk08+UZ06ddSsWbNsX8fhcGjixInq2bOnAgICsnwWlz7ObNhdLleW8fz6nnLLtKb7SqxYsUJvvPGGPvjgAzVt2lR//vmnBg0apNdee00vv/xyts+Ji4vTq6++mmU8Pj5eRYoUkSRFRESoSpUqSkxM1LFjx9zrlC9fXuXLl9fOnTuVkpLiHq9cubIiIyO1efNmnTt3zj1eo0YNhYeHKz4+XtOnT9fXX3+tEydOKCYmRmPHjs3yxTRq1Ejp6elq1aqV4uPjs8TYtGlTjRkzRpIUEhKiwMBAPf300/rtt9/kdDpVqVIljRs3Tu3bt9fBgwe1adMmvf/++1q7dq3OnTunGjVqqF+/fmrYsGG+5HTpL2ZMTIwCAwO1fv36bHNKSEhwjzkcDjVu3FgpKSkeO1RCQkJUt25dJSUlac+ePe7xsLAw1axZUwcPHtSBAwfc4974nsiJnMiJnMgp+5zS0tK0fv16BQcHWyYnK35P5ERO5EROvpRTZGSkHA6H4uPjFRDwf63m/v37Vbp06cvmdPr0aU2dOlX9+/fXmTNndOLEiSzXY6elpWnHjh168cUX3fk6HA6VLl1aR44c8fgMChUqpBMnTqhIkSIe4/n5PWXWyH9j2n2609PTVbhwYc2cOVPdunVzj/fu3VvJycmaM2dOlue0bt1azZo185ga/quvvtKjjz6q06dPZ7u3Ibsj3RUqVNDx48fd91LL771PU6ZMUd++fTV+/Hg1adJE48aN08yZM7VlyxaP0x4y9z4dO3ZM6enp7vGUlBTVrVtXQ4cO1csvvyyHw6E9e/aoefPmevjhh3XfffcpNDRUW7duVbNmzVSmTBm5XC517NhRycnJGjt2rCIjIzVt2jSNGDFCq1evVv369a8qJ1/eo0ZO5ERO5EROWcfHjx+v0aNH6/Dhw6pbt67GjRunBg0aZBt7u3bt9PPPP+ufmjdvruXLl8vhcOjhhx/Wl19+6bH8lltu0YIFC9yxV65cWfv27fNY54033tCzzz6bLzlZ8XsiJ3IiJ3KyWk7NmzdX48aN3QcPM+vDgAED9Oyzz+aY0+TJk/Xkk09q3759KlGihNavX68GDRq485CkRx55RJs3b9bq1as9ctq1a5dq1aql1atXuw84Ll68WJ07d3Y3/FeTU07f0+nTp3N1n24ZJmrSpIkxYMAA92On02mUK1fOiIuLy3b9Bg0aGM8995zH2JQpU4yQkBAjIyMjV++ZkpJiSDJSUlKuPPB/0aRJE+PJJ590P3Y6nUbZsmVzzOuf/ve//xnFihUzli1bZly4cMEwDMPo3r278eCDD172eUWKFDG++OILj7ESJUoYn3zySR4zAAD4s2nTphmBgYHGxIkTjS1bthj9+/c3wsPDjSNHjmS7/vHjx41Dhw65fzZv3mw4HA7jpZdecteh3r17G7feeqvHeidOnPB4nejoaGPkyJEe65w+fdrr+QIAfMe0adOMoKAg4/PPPze2bt1qPProo0Z4eLhx+PBhwzAM46GHHjKGDh2a5XmtWrUyunfvbhiGYVy4cMFYtWqVuwYZxsU+rnDhwsaHH36Y7fveeuutRv369Y01a9YYv/76q1G1alWjZ8+eXsjw/+S2tzR19vLY2Fh98sknmjx5srZt26b//Oc/OnPmjPr27StJ6tWrl4YNG+Ze//bbb9eHH36oadOmKTExUYsXL9bLL7+s22+/3WMPiJm8MWOfy+XS/PnzVa1aNXXs2FGRkZFq2rSpZs+e7fG8Fi1aaPr06Tpx4oRcLpemTZumtLQ0tWvXLr/SAwD4AW/crkWSgoKCPNYrXrx4ltcqVqyYxzqZl3IBAAqG7t27a9SoURo+fLjq1aunjRs3auHChe65vPbv369Dhw55PGfHjh369ddfPWYk/6dp06bJMAz17Nkz2+Vff/21atSoofbt2+u2225Tq1at9PHHH+dfYlfDq61/LowbN8647rrrjMDAQKNJkybG6tWr3cvatm1r9O7d2/34woULxiuvvGJUqVLFCA4ONipUqGA88cQTxsmTJ3P9ft4+0v33338bkoyVK1d6jD/77LNGkyZN/vX5a9asMSQZq1atMv7880/D6XQahw4dMiQZhQsXNkaPHm3Ex8cbcXFxhs1mM1asWOF+7smTJ41bbrnFkGQEBAQYoaGhxqJFi/I9RwCA7zp//rzhcDiM7777zmO8V69eRteuXXP1GnXq1DEeeeQRdx0yjItHusPCwoyIiAijWrVqxuOPP24kJSV5PC86OtqIiooySpQoYdSrV894++23PY5SAACQG06n06MG+arc9pamT6Q2YMAADRgwINtlK1as8HgcEBCgESNGaMSIEdcgMnN89tlnuuGGG9wz9kn/NwvgHXfcocGDB0uS6tWrp5UrV2rChAlq27atJOnll19WcnKylixZolKlSmn27Nm677779Msvv+iGG2649skAAK65K7k7yKUyb9fy2WefqUqVKu7xW2+9VXfddZcqVaqk3bt364UXXlCnTp20atUq99lmAwcOVIMGDVSiRAmtXLlSw4YN06FDhzR69Oj8TRIAYAln0zNyXFamQrTSMlySXFmWFQ40vY3NE/+K1g+UKlVKDodDR44c8Rg/cuSIxwX82Tlz5oymTZumkSNHyuVyKTExUZUqVVKpUqUUEBCgWrVqeaxfs2ZN/frrr5Kk3bt36/3339fmzZtVu3ZtSVLdunX1yy+/aPz48ZowYUI+ZgkAsKrMnb+NGjXS7t27ValSJdntdvXo0cO9zg033KCYmBhVqVJFK1asUPv27SVdvGwsU+Zsuo899pji4uIUFBR0zXMBAPi2WsMXXdHz9r7ZOZ8j8S5Tr+m2osDAQDVs2FBLly51j7lcLi1dulTNmze/7HNnzJih8+fP68EHH5TL5dKxY8fkcrkUGBioxo0ba8eOHR7r79y5U9HR0ZKks2fPSsp6v7jMmfkAAAVDfuz87devn0cdyk7lypVVqlQp/fnnnzm+XtOmTZWRkaG9e/fmOQ8AAKyCI91eEBsbq969e6tRo0Zq0qSJxowZk2WCuHLlyikuLs7jeZ999pm6deumkiVLZrkVwLPPPqvu3burTZs2uvHGG7Vw4ULNnTvXfQp+jRo1dP311+uxxx7TqFGjVLJkSc2ePVuLFy/WvHnzrkneAADzXbrzN/OWnJk7f3O6nCvTpTt//82BAwd0/PhxlSlTJsd1Nm7cKLvd7nG7TAAAMm0d2dHj8fkLTn2/6aBGfL9VktSuWindWqeMOtaOUlAh35g4+0rQdHtB9+7ddezYMQ0fPlyHDx9WvXr1sszY988j0pkz9s1b8INmbTigRZsP6a8jKaqwM14d65TRbV26asKECYqLi9PAgQNVvXp1ffvtt2rVqpWkizd/X7BggYYOHarbb79dp0+f1vXXX6/Jkyfrtttuu+afAQDAPPm98/f06dN69dVXdffdd6t06dLavXu3nnvuOV1//fXq2PHiH0yrVq3SmjVrdOONN6pYsWJatWqVBg8erAcffDDbWc4BALj02uzFW4/omRkblXru/+rPz7uStGJnkt74YZtG31tPHWpFZfcyPs9mGJfc4bsASE1Nzd0NzE1w6YZmt0kuQ+7/Dw0J8OsNDQBwbb3//vt655133Dt/33vvPTVt2lSS1K5dO1WsWFGff/65e/0dO3aoRo0a+vHHH3XzzTfL5XLp4MGDKlu2rM6fP69u3bopPj5eycnJKlu2rG655Ra99tpr7h3KGzZs0BNPPKHt27fr/PnzqlSpkh566CHFxsZyPTcA4LIWbz2iR79cLxlSds2p7f//z8cPNdLNPtQP5ba3pOk2WeaMfcu2H9VTU+IlXWZDkzTu/vq6qUak383YBwAAAAD/lHbBqSZvLNGpcxnZ9kGZbLp4IHLNCx0U7COnmue2t6RzM1luZ+zL3AAH/P/G3N9m7AMA+KacbtfidDr1564/dX3V6923BLsUO38BAPlhwR+HPE4pz4khKeVchn7YfEh31i/v/cDyERUTAIAC7N93/u7NfpSdvwCAfPDjliPuS2r/jd0mLdp8xO+abm4ZZrKtIzvq5lqRstv+fV3p4oZ2cy1mgQUAAADg/5LPpueq4ZYuNubJ59K9G5AXcKTbBySfvZC3De3sBe8GBAAoMArK7VoAAL4pvHBgno50h4cEej+ofEbTbbLcXtN9qXV7T3ohEgBAQVRQbtcCAPBNt9SO0sIth3O1rsuQOtbxvzrE6eUAAMB9u5ZT/5jMJvPIw6lzGer/5Xot3nrEhOgAAFZ12w1lFBoSoH+72tYmKSwkQJ3qlLkWYeUrmm6TbR3ZUfEvd1BocO42tNDgAMW/3OFahAYAKCDSLjj1zIyNOd4fVfr/44Y0ZMZGpV1wXrvgAACWFlzIodH31pNsyrEfyrxP97v31vOZ24XlBU23yQoHBqh4kSCNvq9erja00ffVU/EiQdcuQACA5WXeruXfLqe79HYtAABcrbPpGTqbnqEW15fUuJ71VSzY8+rnzMmmiwUH6P2e9dXi+pI53urSl3FNt4/oUCtKHz/USENmbFTKuQz3ZAKZ/x8aEqB3uZYOAOAFBeF2LQAA3/Nv81tl1qXUtAw9OSXePe5vt62k6fYhN9eK0poXOuiHzYe08I9D+uvoCVWILKFbbyijTnXK+OWpFAAA31cQbtcCAIBZaLp9THAhh+6sX17d6pVTSkqKwsLCZLPl8ibeAABcgYJwuxYAgO/5520rMxmGodSUVIWGhVqiF6Lp9lE2m03h4eFmhwEAKAAKwu1aAAC+59LbVv5TkciS1zAS72IiNR+VkZGhdevWKSPD/yYKAAD4l4JwuxYAgP+wWi9E0+3DnE5uyQIA8L6CcLsWAIB/sVIvRNMNAADcd9EIDcn+di2hIQH65KFG3EUDAIA84ppuAAAKsEvvd9ry+pJaMaSd5iYc1PA5WyVJbaqW0q03lFHHWlEKKuRwr3+56/AAAMD/sRmGkcubhFhDamqqwsLClJKSotDQULPDyZFhGDp37pxCQkIsMWMfAMA3VRw6/4qe52/3SAUA+A9/6YVy21tyerkPCwzkliwAAAAACh4r9UKcG+ajnE6n1q9fr0aNGikggK8JAOAdOd0jNSMjQxs2xKtBg/rUIQDANWW1Xsj/MwAAAFcsp2uzM+xScIBNhQMDLPEHDwAAZuH0cgAAAAAAvISmGwAAAAAAL2H2ch9lGIacTqccDodPz9gHALAm6hAAwCz+UoOYvdwC0tPTzQ4BAFCAUYcAAGaxUg2i6fZRTqdTCQkJcjqdZocCACiAqEMAALNYrQbRdAMAAAAA4CU03QAAAAAAeAlNtw9zOBxmhwAAKMCoQwAAs1ipBjF7OQAAAAAAecTs5X7OMAwlJyergO0TAQD4COoQAMAsVqtBNN0+yul0avv27ZaZsQ8A4F+oQwAAs1itBtF0AwAAAADgJTTdAAAAAAB4CU23j7LZbAoJCZHNZjM7FABAAUQdAgCYxWo1iNnLAQAAAADII2Yv93Mul0tHjx6Vy+UyOxQAQAFEHQIAmMVqNYim20e5XC7t2bPHMhsaAMC/UIcAAGaxWg2i6QYAAAAAwEtougEAAAAA8BKabh9ls9kUFhZmmRn7AAD+hToEADCL1WoQs5cDAAAAAJBHzF7u51wulw4cOGCZyQMAAP6FOgQAMIvVahBNt4+y2oYGAPAv1CEAgFmsVoOuqOn+8ssv1bJlS5UtW1b79u2TJI0ZM0Zz5szJ1+AAAAAAAPBneW66P/zwQ8XGxuq2225TcnKynE6nJCk8PFxjxozJ7/gAAAAAAPBbeW66x40bp08++UQvvviiHA6He7xRo0b6448/8jW4gsxutysiIkJ2O1cAAACuPeoQAMAsVqtBAXl9QmJiourXr59lPCgoSGfOnMmXoHBxQ6tSpYrZYQAACijqEADALFarQXnedVCpUiVt3Lgxy/jChQtVs2bN/IgJujh5wO7duy0zeQAAwL9QhwAAZrFaDcrzke7Y2Fg9+eSTSktLk2EYWrt2raZOnaq4uDh9+umn3oixQHK5XDp27Jiio6Mtc1oFAMB/UIcAAGaxWg3Kc9P9yCOPKCQkRC+99JLOnj2r+++/X2XLltXYsWPVo0cPb8QIAAAAAIBfylPTnZGRoSlTpqhjx4564IEHdPbsWZ0+fVqRkZHeig8AAAAAAL+Vp2P1AQEBevzxx5WWliZJKly4MA23l9jtdpUvX94Sp1MAAPwPdQgAYBar1aA8Z9GkSRPFx8d7IxZcwmobGgDAv1CHAABmsVoNyvM13U888YSeeeYZHThwQA0bNlSRIkU8lsfExORbcAWZ0+nUzp07Va1aNY/7oQMAcC1QhwAAZrFaDcpz0505WdrAgQPdYzabTYZhyGazyel05l90BZhhGEpJSZFhGGaHAgAogKhDAACzWK0G5bnpTkxM9EYcAAAAAABYTp6b7ujoaG/EAQAAAACA5eS56Zak3bt3a8yYMdq2bZskqVatWho0aJCqVKmSr8EVZHa7XZUrV7bM5AEAAP9CHQIAmMVqNSjPWSxatEi1atXS2rVrFRMTo5iYGK1Zs0a1a9fW4sWLvRFjgWS32xUZGWmZDQ0A4F+oQwAAs1itBtmMPF6dXr9+fXXs2FFvvvmmx/jQoUP1448/asOGDfkaYH5LTU1VWFiYUlJSFBoaanY4OXI6ndq8ebPq1KljiRn7AAD+hToEADCLv9Sg3PaWed51sG3bNvXr1y/L+MMPP6ytW7fm9eWQA8MwdO7cOcvM2AcA8C/UIQCAWaxWg/LcdEdERGjjxo1Zxjdu3KjIyMj8iAkAAAAAAEvI80Rq/fv316OPPqo9e/aoRYsWkqTffvtNb731lmJjY/M9QAAAAAAA/FWer+k2DENjxozRu+++q4MHD0qSypYtq2effVYDBw6UzWbzSqD5xV+u6c68IXxYWJjPf6YAAOuhDgEAzOIvNSi3vWWem+5LnTp1SpJUrFixK32Ja85fmm4AAAAAgO/y2kRqiYmJ2rVrl6SLzXZmw71r1y7t3bv3yqJFFhkZGVq3bp0yMjLMDgUAUABRhwAAZrFaDcpz092nTx+tXLkyy/iaNWvUp0+f/IgJ/5/T6TQ7BABAAUYdAgCYxUo1KM9Nd3x8vFq2bJllvFmzZtnOag4AAAAAQEGV56bbZrO5r+W+VEpKiqX2RgAAAAAAcLXy3HS3adNGcXFxHg220+lUXFycWrVqla/BFWQOh0MxMTFyOBxmhwIAKICoQwAAs1itBuX5Pt1vvfWW2rRpo+rVq6t169aSpF9++UWpqalatmxZvgdYkAUGBpodAgCgAKMOAQDMYqUalOcj3bVq1VJCQoLuu+8+HT16VKdOnVKvXr20fft21alTxxsxFkhOp1Pr16/nlH0AgCmoQwAAs1itBuX5SLcklS1bVm+88UZ+xwIAAAAAgKXk+kh3UlKS9u3b5zG2ZcsW9e3bV/fdd5+mTJmS78EBAAAAAODPct10P/XUU3rvvffcj48eParWrVtr3bp1On/+vPr06aMvv/zSK0ECAAAAAOCPct10r169Wl27dnU//uKLL1SiRAlt3LhRc+bM0RtvvKHx48d7JciCyOFwqFGjRpaZsQ8A4F+oQwAAs1itBuW66T58+LAqVqzofrxs2TLdddddCgi4eFl4165dtWvXrnwPsCBLT083OwQAQAFGHQIAmMVKNSjXTXdoaKiSk5Pdj9euXaumTZu6H9tsNp0/fz5fgyvInE6nEhISLDNjHwDAv1CHAABmsVoNynXT3axZM7333ntyuVyaOXOmTp06pZtuusm9fOfOnapQoYJXggQAAAAAwB/luul+7bXX9P333yskJETdu3fXc889p+LFi7uXT5s2TW3bts1zAOPHj1fFihUVHByspk2bau3atZddPzk5WU8++aTKlCmjoKAgVatWTQsWLMjz+wIAAAAA4G25vk93TEyMtm3bpt9++02lS5f2OLVcknr06KFatWrl6c2nT5+u2NhYTZgwQU2bNtWYMWPUsWNH7dixQ5GRkVnWT09P180336zIyEjNnDlT5cqV0759+xQeHp6n9/UXVpk4AADgn6hDAACzWKkG2QzDMMx686ZNm6px48Z6//33JUkul0sVKlTQU089paFDh2ZZf8KECXrnnXe0fft2FSpU6IreMzU1VWFhYUpJSVFoaOhVxQ8AAAAAKJhy21vm+kh3fktPT9fvv/+uYcOGucfsdrs6dOigVatWZfuc77//Xs2bN9eTTz6pOXPmKCIiQvfff7+ef/75HPeEnD9/3mOCt9TUVElSRkaGMjIy3O9rt9vlcrnkcrk84rHb7XI6nbp030RO4w6HQzabzf26l45LyjIRQE7jAQEBcrlcOnnypEJDQ2Wz2WSz2eRwOLLEmNO4L+ZkGIbHODmREzmREzn5bk6GYXjUISvklN04OZETOZETOfleTna7XSkpKSpatKi7BvliTrllWtOdlJQkp9OpqKgoj/GoqCht37492+fs2bNHy5Yt0wMPPKAFCxbozz//1BNPPKELFy5oxIgR2T4nLi5Or776apbx+Ph4FSlSRJIUERGhKlWqKDExUceOHXOvU758eZUvX147d+5USkqKe7xy5cqKjIzU5s2bde7cOfd4jRo1FB4ervj4eI8vPCYmRoGBgVq/fr1HDI0aNVJ6eroSEhLcYw6HQ40bN9bJkye1du1ahYeHy2azKSQkRHXr1lVSUpL27NnjXj8sLEw1a9bUwYMHdeDAAfe4L+aUkpLi8d2SEzmREzmRk+/mdO7cOY86ZIWcrPg9kRM5kRM5WTGn2rVru+O7tOn2tZyCg4OVG6adXn7w4EGVK1dOK1euVPPmzd3jzz33nH766SetWbMmy3OqVaumtLQ0JSYmuvdyjB49Wu+8844OHTqU7ftkd6S7QoUKOn78uPsUAF/c+3ThwgWtX79eDRo0cL+uL+19suIeNXIiJ3IiJ3L6v/GMjAyPOmSFnLIbJydyIidyIiffy8kwjCw1yBdzOn36tG+fXl6qVCk5HA4dOXLEY/zIkSMqXbp0ts8pU6aMChUq5PHB16xZU4cPH1Z6eroCAwOzPCcoKEhBQUFZxgMCAhQQ4Jl+5gf6T5e+X27G//m6VzJ+6Snlly7PKca8jpuVU3bj5EROlxsnJ3IiJ3NyyqkO+XNOOY2TEzlJ5JRTjHkdJydykq4+p4yMjGxrUE7rS+bllBu5PxH9/2vbtq2++OILj9MCrkRgYKAaNmyopUuXusdcLpeWLl3qceT7Ui1bttSff/7psedh586dKlOmTLYNtz/LPKX80tMpAAC4VqhDAACzWK0G5bnprl+/voYMGaLSpUurf//+Wr169RW/eWxsrD755BNNnjxZ27Zt03/+8x+dOXNGffv2lST16tXLY6K1//znPzpx4oQGDRqknTt3av78+XrjjTf05JNPXnEMvsrhcKhu3bpXtUcFAIArRR0CAJjFajUoz033mDFjdPDgQU2aNElHjx5VmzZtVKtWLY0aNSrLqeL/pnv37ho1apSGDx+uevXqaePGjVq4cKF7crX9+/d7XKtdoUIFLVq0SOvWrVNMTIwGDhyoQYMGZXt7MX/ncrl09OhRj6P6AABcK9QhAIBZrFaDrnoitaNHj+rjjz/W66+/LqfTqdtuu00DBw7UTTfdlF8x5it/uU935gQ2jRo1yvH6DAAAvIU6BAAwi7/UoNz2lnk+0n2ptWvXasSIEXr33XcVGRmpYcOGqVSpUurSpYuGDBlyNS8NAAAAAIDfy/Nug6NHj+rLL7/UpEmTtGvXLt1+++2aOnWqOnbs6L7QvU+fPrr11ls1atSofA8YAAAAAAB/keemu3z58qpSpYoefvhh9enTRxEREVnWiYmJUePGjfMlwILKZrMpLCzMMjP2AQD8C3UIAGAWq9WgPF/T/csvv6h169beisfr/OWabgAAAACA7/LaNd0jRoxQcnJytm/oq5On+SOXy6UDBw5YZsY+AIB/oQ4BAMxitRqU56b7p59+Unp6epbxtLQ0/fLLL/kSFKy3oQEA/At1CABgFqvVoFxf052QkCBJMgxDW7du1eHDh93LnE6nFi5cqHLlyuV/hAAAAAAA+KlcN9316tWTzWaTzWbL9jTykJAQjRs3Ll+DAwAAAADAn+W66U5MTJRhGKpcubLWrl3rMWt5YGCgIiMj5XA4vBJkQWS32xURESG7/apupQ4AwBWhDgEAzGK1GpTn2cv9HbOXAwAAAACuVm57y1wd6f7+++/VqVMnFSpUSN9///1l1+3atWveIkW2XC6XEhMTValSJcvs4QEA+A/qEADALFarQblqurt166bDhw8rMjJS3bp1y3E9m80mp9OZX7EVaC6XS8eOHVN0dLQlNjQAgH+hDgEAzGK1GpSrpvvSqdqtMm07AAAAAADelqfdBhcuXFD79u21a9cub8UDAAAAAIBl5KnpLlSokPt+3fAuu92u8uXLW+J0CgCA/6EOAQDMYrUalOcsHnzwQX322WfeiAWXsNqGBgDwL9QhAIBZrFaDcn2f7kwZGRmaOHGilixZooYNG6pIkSIey0ePHp1vwRVkTqdTO3fuVLVq1bj/OQDgmqMOAQDMYrUalOeme/PmzWrQoIEkaefOnR7LbDZb/kQFGYahlJQUFbDbqAMAfAR1CABgFqvVoDw33cuXL/dGHAAAAAAAWI41TpIHAAAAAMAH5flItyStX79e33zzjfbv36/09HSPZbNmzcqXwAo6u92uypUrW2byAACAf6EOAQDMYrUalOcspk2bphYtWmjbtm367rvvdOHCBW3ZskXLli1TWFiYN2IskOx2uyIjIy2zoQEA/At1CABgFqvVoDxn8cYbb+h///uf5s6dq8DAQI0dO1bbt2/Xfffdp+uuu84bMRZITqdTmzZtktPpNDsUAEABRB0CAJjFajUoz0337t271blzZ0lSYGCgzpw5I5vNpsGDB+vjjz/O9wALKsMwdO7cOcvM2AcA8C/UIQCAWaxWg/LcdBcvXlynTp2SJJUrV06bN2+WJCUnJ+vs2bP5Gx0AAAAAAH4szxOptWnTRosXL9YNN9yge++9V4MGDdKyZcu0ePFitW/f3hsxAgAAAADgl2xGHo/ZnzhxQmlpaSpbtqxcLpfefvttrVy5UlWrVtVLL72k4sWLeyvWfJGamqqwsDClpKQoNDTU7HBylHlD+LCwMNlsNrPDAQAUMNQhAIBZ/KUG5ba3zHPT7e/8pekGAAAAAPiu3PaWubqmOzU1Ndc/yB8ZGRlat26dMjIyzA4FAFAAUYcAAGaxWg3K1TXd4eHh/3pY3zAM2Ww2y0zr7gv4LAEAZqIOAQDMYqUalKume/ny5d6OAwAAAAAAy8lV0922bVtvxwEAAAAAgOXkqulOSEhQnTp1ZLfblZCQcNl1Y2Ji8iWwgs7hcCgmJkYOh8PsUAAABRB1CABgFqvVoFw13fXq1dPhw4cVGRmpevXqyWazKbtJz7mmO38FBgaaHQIAoACjDgEAzGKlGpSrpjsxMVERERHu/4b3OZ1OrV+/Xo0aNVJAQK6+JgAA8g11CABgFqvVoFxlEB0dne1/AwAAAACAnF3RboODBw/q119/1dGjR+VyuTyWDRw4MF8CAwAAAADA3+W56f7888/12GOPKTAwUCVLlvS4f7fNZqPpBgAAAADg/7MZ2c2IdhkVKlTQ448/rmHDhslut3srLq9JTU1VWFiYUlJSFBoaanY4OTIMQ06nUw6Hw2PHBgAA1wJ1CABgFn+pQbntLfPcNZ89e1Y9evTwy4bb36Snp5sdAgCgAKMOAQDMYqUalOfOuV+/fpoxY4Y3YsElnE6nEhISuAUbAMAU1CEAgFmsVoPyfE13XFycunTpooULF+qGG25QoUKFPJaPHj0634IDAAAAAMCfXVHTvWjRIlWvXl2SskykBgAAAAAALspz0/3uu+9q4sSJ6tOnjxfCwaUcDofZIQAACjDqEADALFaqQXluuoOCgtSyZUtvxIJLBAQEqHHjxmaHAQAooKhDAACzWK0G5XkitUGDBmncuHHeiAWXMAxDycnJyuMd3QAAyBfUIQCAWaxWg/J8pHvt2rVatmyZ5s2bp9q1a2eZSG3WrFn5FlxB5nQ6tX37djVq1EgBAXn+mgAAuCrUIQCAWaxWg/KcQXh4uO666y5vxAIAAAAAgKXkuemeNGmSN+IAAAAAAMBy8nxNN64Nm82mkJAQbsMGADAFdQgAYBar1SCbkYur0xs0aKClS5eqePHiql+//mWT37BhQ74GmN9SU1MVFhamlJQUhYaGmh0OAAAAAMAP5ba3zNXp5XfccYeCgoIkSd26dcuXAHF5LpdLSUlJKlWqlOx2TkgAAFxb1CEAgFmsVoNy1XSPGDEi2/+G97hcLu3Zs0clSpSwxIYGAPAv1CEAgFmsVoOuav71tLQ0TZ8+XWfOnNHNN9+sqlWr5ldcAAAAAAD4vVw33bGxsbpw4YLGjRsnSUpPT1ezZs20detWFS5cWM8995x+/PFHtWjRwmvBAgAAAADgT3J9rP7HH3/UzTff7H789ddfa//+/dq1a5dOnjype++9V6+//rpXgiyIbDabwsLCLDNjHwDAv1CHAABmsVoNynXTvX//ftWqVcv9+Mcff9Q999yj6Oho2Ww2DRo0SPHx8V4JsiByOByqWbOmHA6H2aEAAAog6hAAwCxWq0G5brrtdrsuvbvY6tWr1axZM/fj8PBwnTx5Mn+jK8BcLpcOHDggl8tldigAgAKIOgQAMIvValCum+6aNWtq7ty5kqQtW7Zo//79uvHGG93L9+3bp6ioqPyPsICy2oYGAPAv1CEAgFmsVoNyPZHac889px49emj+/PnasmWLbrvtNlWqVMm9fMGCBWrSpIlXggQAAAAAwB/l+kj3nXfeqQULFigmJkaDBw/W9OnTPZYXLlxYTzzxRL4HCAAAAACAv7IZl16oXQCkpqYqLCxMKSkpCg0NNTucHLlcLiUmJqpSpUqWuCE8AMC/UIcAAGbxlxqU294y16eX49qy2+2qUqWK2WEAAAoo6hAAwCxWq0G+u9uggHO5XNq9e7dlJg8AAPgX6hAAwCxWq0E03T7K5XLp2LFjltnQAAD+hToEADCL1WoQTTcAAAAAAF6S56Y7MTFRu3btyjK+a9cu7d27Nz9iAgAAAADAEvLcdPfp00crV67MMr5mzRr16dMnP2KCLk4eUL58eZ+erQ8AYF3UIQCAWaxWg/KcRXx8vFq2bJllvFmzZtq4cWN+xARZb0MDAPgX6hAAwCxWq0F5zsJms+nUqVNZxlNSUuR0OvMlKEhOp1Pbtm3jMwUAmII6BAAwi9VqUJ6b7jZt2iguLs7jA3A6nYqLi1OrVq3yNbiCzDAMpaSkyDAMs0MBABRA1CEAgFmsVoMC8vqEt956S23atFH16tXVunVrSdIvv/yi1NRULVu2LN8DBAAAAADAX+X5SHetWrWUkJCg++67T0ePHtWpU6fUq1cvbd++XXXq1PFGjAAAAAAA+CWbYZVj9rmUmpqqsLAwpaSkKDQ01OxwcuRyuZSUlKRSpUpZZgIBAID/oA4BAMziLzUot71lrk4vT0hIUJ06dWS325WQkHDZdWNiYvIWKbJlt9sVGRlpdhgAgAKKOgQAMIvValCumu569erp8OHDioyMVL169WSz2bK9qN1ms1lmhjmzOZ1Obd68WXXq1JHD4TA7HABAAUMdAgCYxWo1KFdNd2JioiIiItz/De8zDEPnzp2zzIx9AAD/Qh0CAJjFajUoV013dHS0+7/37dunFi1aKCDA86kZGRlauXKlx7oAAAAAABRkeb4q/cYbb9SJEyeyjKekpOjGG2/Ml6AAAAAAALCCPDfdhmHIZrNlGT9+/LiKFCmSL0FBcjgcqlGjhiWuYQAA+B/qEADALFarQbk6vVyS7rrrLkkXJ0vr06ePgoKC3MucTqcSEhLUokWL/I+wgLLZbAoPDzc7DABAAUUdAgCYxWo1KNdHusPCwhQWFibDMFSsWDH347CwMJUuXVqPPvqovvrqK2/GWqBkZGRo3bp1ysjIMDsUAEABRB0CAJjFajUo10e6J02aJEmqWLGihgwZwqnk1wC3XwMAmIk6BAAwi5VqUK6b7kwjRozwRhwAAAAAAFhOnidSO3LkiB566CGVLVtWAQEBcjgcHj9XYvz48apYsaKCg4PVtGlTrV27NlfPmzZtmmw2m7p163ZF7wsAAAAAgDfl+Uh3nz59tH//fr388ssqU6ZMtjOZ58X06dMVGxurCRMmqGnTphozZow6duyoHTt2KDIyMsfn7d27V0OGDFHr1q2v6v19lcPhUExMjGVm7AMA+BfqEADALFarQTbDMIy8PKFYsWL65ZdfVK9evXwJoGnTpmrcuLHef/99SZLL5VKFChX01FNPaejQodk+x+l0qk2bNnr44Yf1yy+/KDk5WbNnz87V+6WmpiosLEwpKSkKDQ3Nlxy8wTAMOZ1OORyOq96xAQBAXlGHAABm8ZcalNveMs+nl1eoUEF57NNzlJ6ert9//10dOnT4v4DsdnXo0EGrVq3K8XkjR45UZGSk+vXrly9x+CKn06n169dbagIBAID/oA4BAMxitRqU59PLx4wZo6FDh+qjjz5SxYoVr+rNk5KS5HQ6FRUV5TEeFRWl7du3Z/ucX3/9VZ999pk2btyYq/c4f/68zp8/736cmpoq6eI09JlT0NvtdtntdrlcLrlcLve6meNOp9NjR0NO45l7Yv45tX3maRH/3GhyGg8ICJBhGO49PNLFe9U5HI4sMeY07qs5XTpOTuRETuRETr6b0z/rkBVyym6cnMiJnMiJnHwvp+xqkC/mlFt5brq7d++us2fPqkqVKipcuLAKFSrksfzEiRN5fclcO3XqlB566CF98sknKlWqVK6eExcXp1dffTXLeHx8vPu2ZxEREapSpYoSExN17Ngx9zrly5dX+fLltXPnTqWkpLjHK1eurMjISG3evFnnzp1zj9eoUUPh4eGKj4/3+MJjYmIUGBio9evXe8TQqFEjpaenKyEhwT3mcDjUuHFjpaamKjk5WRs2bJDNZlNISIjq1q2rpKQk7dmzx71+WFiYatasqYMHD+rAgQPucV/MKSUlxWNnCjmREzmREzn5bk5paWkedcgKOVnxeyInciIncrJiTrVr15Ykdw3y1ZyCg4OVG3m+pnvy5MmXXd67d+9cv1Z6eroKFy6smTNnesxA3rt3byUnJ2vOnDke62/cuFH169f3uKA+cy+E3W7Xjh07VKVKFY/nZHeku0KFCjp+/Lj7vHtf3Pt04cIFrV+/Xg0aNHC/ri/tfbLiHjVyIidyIidy+r/xjIwMjzpkhZyyGycnciInciIn38vJMIwsNcgXczp9+nSurunOc9Od35o2baomTZpo3Lhxki420dddd50GDBiQZSK1tLQ0/fnnnx5jL730kk6dOqWxY8eqWrVqCgwMvOz7MZEaAAD/jjoEADCLv9Sg3PaWeT69/FJpaWlKT0/3GMtrIxsbG6vevXurUaNGatKkicaMGaMzZ86ob9++kqRevXqpXLlyiouLU3BwsOrUqePx/PDwcEnKMm4F6enpCgkJMTsMAEABRR0CAJjFSjUoz7OXnzlzRgMGDFBkZKSKFCmi4sWLe/zkVffu3TVq1CgNHz5c9erV08aNG7Vw4UL35Gr79+/XoUOH8vy6/s7pdCohISHLqRMAAFwL1CEAgFmsVoPyfKT7ueee0/Lly/Xhhx/qoYce0vjx4/X333/ro48+0ptvvnlFQQwYMEADBgzIdtmKFSsu+9zPP//8it4TAAAAAABvy3PTPXfuXH3xxRdq166d+vbtq9atW+v6669XdHS0vv76az3wwAPeiBMAAAAAAL+T59PLT5w4ocqVK0u6eP125i3CWrVqpZ9//jl/oyvgLp2pDwCAa406BAAwi5VqUJ6b7sqVKysxMVHSxfuqffPNN5IuHgHPnNQMVy8gIECNGzdWQMBVzXUHAMAVoQ4BAMxitRqU56a7b9++2rRpkyRp6NChGj9+vIKDgzV48GA9++yz+R5gQWUYhpKTk2XyHd0AAAUUdQgAYBar1aCrvk/3vn379Pvvv+v6669XTExMfsXlNf5yn+6MjAytX79ejRo1ssweHgCA/6AOAQDM4i816Jrcp1uSoqOjFR0dfbUvAwAAAACA5VxR071u3TotX75cR48elcvl8lg2evTofAkMAAAAAAB/l+em+4033tBLL72k6tWrKyoqSjabzb3s0v/G1bHZbAoJCeEzBQCYgjoEADCL1WpQnq/pjoqK0ltvvaU+ffp4KSTv8pdrugEAAAAAviu3vWWeZy+32+1q2bLlVQWHf+dyubI9fR8AgGuBOgQAMIvValCem+7Bgwdr/Pjx3ogFl3C5XNqzZ49lNjQAgH+hDgEAzGK1GpTna7qHDBmizp07q0qVKqpVq5YKFSrksXzWrFn5FhwAAAAAAP4sz033wIEDtXz5ct14440qWbKkZS5uBwAAAAAgv+W56Z48ebK+/fZbde7c2Rvx4P+z2WwKCwtjpwYAwBTUIQCAWaxWg/LcdJcoUUJVqlTxRiy4hMPhUM2aNc0OAwBQQFGHAABmsVoNyvNEaq+88opGjBihs2fPeiMe/H8ul0sHDhywzOQBAAD/Qh0CAJjFajUoz0e633vvPe3evVtRUVGqWLFilonUNmzYkG/BFWSZG1rp0qVlt+d53wgAAFeFOgQAMIvValCem+5u3bp5IQwAAAAAAKwnz033iBEjvBEHAAAAAACWc0XH6pOTk/Xpp59q2LBhOnHihKSLp5X//fff+RpcQWa32xUREWGJ0ykAAP6HOgQAMIvValCej3QnJCSoQ4cOCgsL0969e9W/f3+VKFFCs2bN0v79+/XFF194I84Cx263M0s8AMA01CEAgFmsVoPyvOsgNjZWffr00a5duxQcHOwev+222/Tzzz/na3AFmcvl0u7duy0zYx8AwL9QhwAAZrFaDcpz071u3To99thjWcbLlSunw4cP50tQuLihHTt2zDIbGgDAv1CHAABmsVoNynPTHRQUpNTU1CzjO3fuVERERL4EBQAAAACAFeS56e7atatGjhypCxcuSJJsNpv279+v559/XnfffXe+BwgAAAAAgL/Kc9P97rvv6vTp04qMjNS5c+fUtm1bXX/99SpWrJhef/11b8RYINntdpUvX94yM/YBAPwLdQgAYBar1SCbYRjGlTzxt99+06ZNm3T69Gk1aNBAHTp0yO/YvCI1NVVhYWFKSUlRaGio2eEAAAAAAPxQbnvLK266/ZW/NN1Op1M7d+5UtWrV5HA4zA4HAFDAUIcAAGbxlxqU294y18frV61apXnz5nmMffHFF6pUqZIiIyP16KOP6vz581ceMTwYhqGUlBQVsH0iAAAfQR0CAJjFajUo1033yJEjtWXLFvfjP/74Q/369VOHDh00dOhQzZ07V3FxcV4JEgAAAAAAf5Trpnvjxo1q3769+/G0adPUtGlTffLJJ4qNjdV7772nb775xitBAgAAAADgj3LddJ88eVJRUVHuxz/99JM6derkfty4cWP99ddf+RtdAWa321W5cmXLzNgHAPAv1CEAgFmsVoNynUVUVJQSExMlSenp6dqwYYOaNWvmXn7q1CkVKlQo/yMsoOx2uyIjIy2zoQEA/At1CABgFqvVoFxncdttt2no0KH65ZdfNGzYMBUuXFitW7d2L09ISFCVKlW8EmRB5HQ6tWnTJjmdTrNDAQAUQNQhAIBZrFaDAnK74muvvaa77rpLbdu2VdGiRTV58mQFBga6l0+cOFG33HKLV4IsiAzD0Llz5ywzYx8AwL9QhwAAZrFaDcp1012qVCn9/PPPSklJUdGiRbPcL23GjBkqWrRovgcIAAAAAIC/ynXTnSksLCzb8RIlSlx1MAAAAAAAWIk1rky3IIfDoRo1amQ5owAAgGuBOgQAMIvValCej3Tj2rDZbAoPDzc7DABAAUUdAgCYxWo1iCPdPiojI0Pr1q1TRkaG2aEAAAog6hAAwCxWq0E03T7MKlPkAwD8E3UIAGAWK9Ugmm4AAAAAALyEphsAAAAAAC+h6fZRDodDMTExlpmxDwDgX6hDAACzWK0G0XT7sMDAQLNDAAAUYNQhAIBZrFSDaLp9lNPp1Pr16y01gQAAwH9QhwAAZrFaDaLpBgAAAADAS2i6AQAAAADwEppuAAAAAAC8xGYYhmF2ENdSamqqwsLClJKSotDQULPDyZFhGHI6nXI4HLLZbGaHAwAoYKhDAACz+EsNym1vyZFuH5aenm52CACAAow6BAAwi5VqEE23j3I6nUpISLDMjH0AAP9CHQIAmMVqNYimGwAAAAAAL6HpBgAAAADAS2i6fZjD4TA7BABAAUYdAgCYxUo1iNnLAQAAAADII2Yv93OGYSg5OVkFbJ8IAMBHUIcAAGaxWg2i6fZRTqdT27dvt8yMfQAA/0IdAgCYxWo1iKYbAAAAAAAvoekGAAAAAMBLaLp9lM1mU0hIiGw2m9mhAAAKIOoQAMAsVqtBzF4OAAAAAEAeMXu5n3O5XDp69KhcLpfZoQAACiDqEADALFarQTTdPsrlcmnPnj2W2dAAAP6FOgQAMIvVahBNNwAAAAAAXkLTDQAAAACAl9B0+yibzaawsDDLzNgHAPAv1CEAgFmsVoOYvRwAAAAAgDxi9nI/53K5dODAActMHgAA8C/UIQCAWaxWg2i6fZTVNjQAgH+hDgEAzGK1GkTTDQAAAACAl9B0AwAAAADgJTTdPsputysiIkJ2O18RAODaow4BAMxitRoUYHYAyJ7dbleVKlXMDgMAUEBRhwAAZrFaDbLGrgMLcrlc2r17t2UmDwAA+BfqEADALFarQTTdPsrlcunYsWOW2dAAAP6FOgQAMIvVahBNNwAAAAAAXkLTDQAAAACAl9B0+yi73a7y5ctbZsY+AIB/oQ4BAMxitRrE7OU+KnNDAwDADNQhAIBZrFaDrLHrwIKcTqe2bdsmp9NpdigAgAKIOgQAMIvVahBNt48yDEMpKSkyDMPsUAAABRB1CABgFqvVIJpuAAAAAAC8hKYbAAAAAAAvoen2UXa7XZUrV7bMjH0AAP9CHQIAmMVqNcgnshg/frwqVqyo4OBgNW3aVGvXrs1x3U8++UStW7dW8eLFVbx4cXXo0OGy6/sru92uyMhIy2xoAAD/Qh0CAJjFajXI9CymT5+u2NhYjRgxQhs2bFDdunXVsWNHHT16NNv1V6xYoZ49e2r58uVatWqVKlSooFtuuUV///33NY7cu5xOpzZt2mSZGfsAAP6FOgQAMIvVapDpTffo0aPVv39/9e3bV7Vq1dKECRNUuHBhTZw4Mdv1v/76az3xxBOqV6+eatSooU8//VQul0tLly69xpF7l2EYOnfunGVm7AMA+BfqEADALFarQQFmvnl6erp+//13DRs2zD1mt9vVoUMHrVq1KlevcfbsWV24cEElSpTIdvn58+d1/vx59+PU1FRJUkZGhjIyMtzvabfb5XK55HK5PGKx2+1yOp0eX3hO4w6HQzabzf26l45LyrKnJqfxgIAAGYYhwzDcy2w2mxwOR5YYcxr31ZwuHScnciInciIn383pn3XICjllN05O5ERO5EROvpdTdjXIF3PKLVOb7qSkJDmdTkVFRXmMR0VFafv27bl6jeeff15ly5ZVhw4dsl0eFxenV199Nct4fHy8ihQpIkmKiIhQlSpVlJiYqGPHjrnXKV++vMqXL6+dO3cqJSXFPV65cmVFRkZq8+bNOnfunHu8Ro0aCg8PV3x8vMcXHhMTo8DAQK1fv94jhkaNGik9PV0JCQnuMYfDocaNGys1NVXJycnasGGDbDabQkJCVLduXSUlJWnPnj3u9cPCwlSzZk0dPHhQBw4ccI/7Yk4pKSke3ys5kRM5kRM5+W5OaWlpHnXICjlZ8XsiJ3IiJ3KyYk61a9eWJHcN8tWcgoODlRs2w8Rj9gcPHlS5cuW0cuVKNW/e3D3+3HPP6aefftKaNWsu+/w333xTb7/9tlasWKGYmJhs18nuSHeFChV0/PhxhYaGSvLNvU8ul0snT55UaGiobDabz+19suIeNXIiJ3IiJ3L6v3HDMDzqkBVyym6cnMiJnMiJnHwvJ7vdrpSUFBUtWtSj6fa1nE6fPq2wsDClpKS4e8vsmNp0p6enq3Dhwpo5c6a6devmHu/du7eSk5M1Z86cHJ87atQo/fe//9WSJUvUqFGjXL9nampqrj4YAAAAAAByktve0tSJ1AIDA9WwYUOPSdBcrouTol165Puf3n77bb322mtauHBhnhpuf5KRkaF169Zl2esDAMC1QB0CAJjFajXI1Gu6JSk2Nla9e/dWo0aN1KRJE40ZM0ZnzpxR3759JUm9evVSuXLlFBcXJ0l66623NHz4cE2ZMkUVK1bU4cOHJUlFixZV0aJFTcvDG/552gQAANcSdQgAYBYr1SDTm+7u3bvr2LFjGj58uA4fPqx69epp4cKF7snV9u/f7zEz3Icffqj09HTdc889Hq8zYsQIvfLKK9cydAAAAAAALsv0pluSBgwYoAEDBmS7bMWKFR6P9+7d6/2AAAAAAADIB6Ze042cORwOxcTEuGfoAwDgWqIOAQDMYrUaRNPtwwIDA80OAQBQgFGHAABmsVINoun2UU6nU+vXr7fUBAIAAP9BHQIAmMVqNYimGwAAAAAAL6HpBgAAAADAS2i6AQAAAADwEpthGIbZQVxLqampCgsLU0pKikJDQ80OJ0eGYcjpdMrhcMhms5kdDgCggKEOAQDM4i81KLe9JUe6fVh6errZIQAACjDqEADALFaqQTTdPsrpdCohIcEyM/YBAPwLdQgAYBar1SCabgAAAAAAvISmGwAAAAAAL6Hp9mEOh8PsEAAABRh1CABgFivVIGYvBwAAAAAgj5i93M8ZhqHk5GQVsH0iAAAfQR0CAJjFajWIpttHOZ1Obd++3TIz9gEA/At1CABgFqvVIJpuAAAAAAC8hKYbAAAAAAAvoen2UTabTSEhIbLZbGaHAgAogKhDAACzWK0GMXs5AAAAAAB5xOzlfs7lcuno0aNyuVxmhwIAKICoQwAAs1itBtF0+yiXy6U9e/ZYZkMDAPgX6hAAwCxWq0E03QAAAAAAeAlNNwAAAAAAXkLT7aNsNpvCwsIsM2MfAMC/UIcAAGaxWg1i9nIAAAAAAPKI2cv9nMvl0oEDBywzeQAAwL9QhwAAZrFaDaLp9lFW29AAAP6FOgQAMIvVahBNNwAAAAAAXkLTDQAAAACAl9B0+yi73a6IiAjZ7XxFAIBrjzoEADCL1WpQgNkBIHt2u11VqlQxOwwAQAFFHQIAmMVqNcgauw4syOVyaffu3ZaZPAAA4F+oQwAAs1itBtF0+yiXy6Vjx45ZZkMDAPgX6hAAwCxWq0E03QAAAAAAeAlNNwAAAAAAXkLT7aPsdrvKly9vmRn7AAD+hToEADCL1WoQs5f7qMwNDQAAM1CHAABmsVoNssauAwtyOp3atm2bnE6n2aEAAAog6hAAwCxWq0E03T7KMAylpKTIMAyzQwEAFEDUIQCAWaxWg2i6AQAAAADwEppuAAAAAAC8hKbbR9ntdlWuXNkyM/YBAPwLdQgAYBar1SBmL/dRdrtdkZGRZocBACigqEMAALNYrQZZY9eBBTmdTm3atMkyM/YBAPwLdQgAYBar1SCabh9lGIbOnTtnmRn7AAD+hToEADCL1WoQTTcAAAAAAF5C0w0AAAAAgJfQdPsoh8OhGjVqyOFwmB0KAKAAog4BAMxitRrE7OU+ymazKTw83OwwAAAFFHUIAGAWq9UgjnT7qIyMDK1bt04ZGRlmhwIAKICoQwAAs1itBtF0+zCrTJEPAPBP1CEAgFmsVINougEAAAAA8BKabgAAAAAAvISm20c5HA7FxMRYZsY+AIB/oQ4BAMxitRpE0+3DAgMDzQ4BAFCAUYcAAGaxUg2i6fZRTqdT69evt9QEAgAA/0EdAgCYxWo1iKYbAAAAAAAvoekGAAAAAMBLaLoBAAAAAPASm2EYhtlBXEupqakKCwtTSkqKQkNDzQ4nR4ZhyOl0yuFwyGazmR0OAKCAoQ4BAMziLzUot70lR7p9WHp6utkhAAAKMOoQAMAsVqpBNN0+yul0KiEhwTIz9gEA/At1CABgFqvVIJpuAAAAAAC8hKYbAAAAAAAvoen2YQ6Hw+wQAAAFGHUIAGAWK9UgZi8HAAAAACCPmL3czxmGoeTkZBWwfSIAAB9BHQIAmMVqNYim20c5nU5t377dMjP2AQD8C3UIAGAWq9Ugmm4AAAAAALyEphsAAAAAAC+h6fZRNptNISEhstlsZocCACiAqEMAALNYrQYxezkAAAAAAHnE7OV+zuVy6ejRo3K5XGaHAgAogKhDAACzWK0G0XT7KJfLpT179lhmQwMA+BfqEADALFarQTTdAAAAAAB4CU03AAAAAABeQtPto2w2m8LCwiwzYx8AwL9QhwAAZrFaDWL2cgAAAAAA8ojZy/2cy+XSgQMHLDN5AADAv1CHAABmsVoNoun2UVbb0AAA/oU6BAAwi9VqEE03AAAAAABeQtMNAAAAAICX0HT7KLvdroiICNntfEUAgGuPOgQAMIvValCA2QEge3a7XVWqVDE7DABAAUUdAgCYxWo1yCd2HYwfP14VK1ZUcHCwmjZtqrVr1152/RkzZqhGjRoKDg7WDTfcoAULFlyjSK8dl8ul3bt3W2byAACAf6EOAQDMYrUaZHrTPX36dMXGxmrEiBHasGGD6tatq44dO+ro0aPZrr9y5Ur17NlT/fr1U3x8vLp166Zu3bpp8+bN1zhy73K5XDp27JhlNjQAgH+hDgEAzGK1GmR60z169Gj1799fffv2Va1atTRhwgQVLlxYEydOzHb9sWPH6tZbb9Wzzz6rmjVr6rXXXlODBg30/vvvX+PIAQAAAAC4PFOv6U5PT9fvv/+uYcOGucfsdrs6dOigVatWZfucVatWKTY21mOsY8eOmj17drbrnz9/XufPn3c/TklJkSSdOHFCGRkZ7ve02+1yuVwee1Myx51OpwzD+Ndxh8Mhm83mft1LxyXJ6XTmajwgIEAXLlzQ6dOndfLkSffrOhyOLDHmNO6LORmG4TFOTuRETuRETr6bU0ZGhkcdskJO2Y2TEzmREzmRk+/lZBhGlhrkizmdPn3aHe/lmNp0JyUlyel0KioqymM8KipK27dvz/Y5hw8fznb9w4cPZ7t+XFycXn311SzjlSpVusKoAQAAAAC46NSpUwoLC8txueVnLx82bJjHkXGXy6UTJ06oZMmSstlsJkZ2eampqapQoYL++usvhYaGmh0OAKCAoQ4BAMziLzXIMAydOnVKZcuWvex6pjbdpUqVksPh0JEjRzzGjxw5otKlS2f7nNKlS+dp/aCgIAUFBXmMhYeHX3nQ11hoaKhPb2gAAGujDgEAzOIPNehyR7gzmTqRWmBgoBo2bKilS5e6x1wul5YuXarmzZtn+5zmzZt7rC9JixcvznF9AAAAAADMYvrp5bGxserdu7caNWqkJk2aaMyYMTpz5oz69u0rSerVq5fKlSunuLg4SdKgQYPUtm1bvfvuu+rcubOmTZum9evX6+OPPzYzDQAAAAAAsjC96e7evbuOHTum4cOH6/Dhw6pXr54WLlzonixt//79stv/74B8ixYtNGXKFL300kt64YUXVLVqVc2ePVt16tQxKwWvCAoK0ogRI7KcGg8AwLVAHQIAmMVqNchm/Nv85gAAAAAA4IqYek03AAAAAABWRtMNAAAAAICX0HQDAAAAAOAlNN0AAAAAAHgJTTcAAAAAAF5C020BTEAPALjWMmuPy+UyORIAQEH1ySefmB1CrtB0W4DNZtPUqVP13nvv0YADALwqs8k+efKkJMlu508JAMC1t2rVKj322GN66qmnzA7lX1Ep/Vhmg71v3z498sgjki424AAAeIPL5ZLdbld8fLxKlSqldevWmR0SAKCAatSokb755htNmjRJTzzxhNnhXFaA2QHgytlsNv3888/atWuXnnrqKQ0cONDskAAAFpXZcG/cuFFt27ZVbGysGjdunGU9wzDYAQwA8CqXy6VChQrpnnvuUUBAgO677z6VKFFC//3vf80OLVsc6fZjp06d0ujRo9W/f39t2bJFEtfWAQC8w263a/PmzWrRooUGDx6sUaNGyTAMHTp0SPHx8UpPT5d0cYcwlzoBALzFMAz3pU1vv/22li5dqqJFi+qNN97Q4MGDTY4uezaDyujXVq9erbFjx2ru3LlavXq16tSpI6fTKYfDYXZoAAALOXPmjLp166bVq1fr1KlTkqQ777xTe/fu1aZNm9SqVSvdcccdio2N5Ug3AMDrRo4cqXHjxunzzz+X0+nUhg0b9Oabb6p///4aN26c2eF54PRyP5J5yp7L5ZLT6VShQoXUrFkzhYaG6vjx47rlllu0ZMkS1apVi8YbAJCvAgIC9Mgjj+jAgQO67bbb5HQ6FRQUpOHDh6t8+fL64IMPNH36dJUoUUJ9+/Y1O1wAgIXs27dP0dHR7sfnzp3Tb7/9pmeeeUadO3eWJHXo0EGVK1fWww8/rMKFC+utt94yK9wsONLtJzIb7kWLFunLL7/U33//rfr166t3796qW7euduzYodjYWCUkJGjx4sWqUaMGjTcA4Ipld212WlqaFi5cqOeee06lSpXSt99+qzJlyki6eMlT165dFRERoW+++caMkAEAFvT444/r77//1ty5c91j586dU7169XTrrbdq7Nix7vEzZ86oT58++vbbb9WvXz+fuaUY13T7CZvNpu+//15du3ZVcHCw6tevr++++04DBw7UzJkzVb16db311ltq2LChGjZsqJ07d9JwAwCuiMvlks1mU0pKivbs2aO///5bZ86cUXBwsG655RaNHTtWI0aMUGRkpCTJ6XSqWLFiql+/vg4dOsT8IgCAfPPOO+9o1qxZkqTjx49LkkJCQtS7d2/99ttv+vXXX93rFilSRDVr1tTNN9+s/fv3+0w9oun2A4Zh6Pjx44qLi9PIkSP16aefavTo0Vq9erXCwsI0btw4bd++XXXq1NELL7ygrl27cj0dAOCKZM5SvnnzZnXq1Em33XabWrdurWHDhunQoUMqXLiw2rdvr/bt27t37mb+/+HDh1W3bl1qEAAgX2Tu1C1UqJCmTp2qKlWqaOvWrZKkdu3aKTQ0VOPHj9dPP/0kSUpJSdHGjRt13333adGiRbLb7T7ReNN0+yjDMNwbiM1mU5EiRXTq1CkVL15cknThwgVFRUXp008/1e7du/X1119Lkpo0aaLPP/9cVatWNS12AIB/ypwRdseOHbrpppvUvHlzTZw4UY888ohWr16t3377TZIUGBiogID/mxbm7NmzeuGFF7R8+XINGDCAphsAcNUMw3Dv1F2yZInatm2rOnXq6K677tL27dvVokULPfPMMzp58qTuv/9+NWnSRC1atNDevXvVu3dv92tkznRuJq7p9kGnTp1SsWLFJElr1qxRQECAatasqcaNG+u2227TO++84zGZWp8+fXT27FmuoQMAXLXU1FT17t1bkZGR+uijj9zjnTt3VkBAgObMmeOx/pw5czRt2jT9/PPPmjdvnurXr3+tQwYAWMyl84qMGDFCM2bM0Lx581SiRAl16dJFhw8f1rx581SjRg3t2rVLO3fu1IoVK1S6dGkNGjRIAQEBPjW/lfltPzwcOXJEMTExWrZsmX788Ue1bdtWp06dUuHChfXCCy9o9OjRmjhxoux2uwoVKiRJOnnypKKiokyOHABgBcnJySpZsqRuv/12SRfPrJKk22+/XU6nU5I87sPdqFEj1a1bVz/99BMNNwAgX2Q23Js3b9bGjRv1wQcfqHLlygoPD9e8efNUunRpdenSRdu3b1fVqlXVuXNnvfPOO3rmmWcUEBCgjIwMn2m4JW4Z5nPS0tLUtWtX3XnnnTp//rymTp2qdu3ayTAMde3aVS+//LIeeeQRbdiwQRUqVNCBAwe0bNkyrVmzxuzQAQAWcN1116lnz55q3769pP+7XrtIkSI6c+aMx7onT55UuXLl9Nxzz/nE6XsAAOv48MMP9dVXX8nlcql69eqSLs47ktl433777erWrZtmzJihG264weO5l14C5QuokD4mOjpaN910k06dOiW73a4iRYq4lxUrVkzPPfecpk+frrVr12r27NnatWuXfvvtN9WqVcvEqAEAVpA5l0hmw505qZoknT59WsePH3fPbP7aa6/pkUce0YULF7iGGwCQ7+rVq6cjR45o06ZNWrt2rSS5J0YLDw/X3LlzZRiGXn/9dZMj/Xdc0+0jMq9bSE1N1d69e5WYmKiffvpJn332mT7//HPdeeed7j90bDabLly4oEKFCuns2bMqXLiw2eEDAPxYZg06f/68JCkoKCjLsilTpmj8+PH67bff9PLLL+utt97SmjVrOKUcAHBVDMPIccKz+Ph4de/eXdWqVdPLL7+spk2bSvq/ncKnT59WSEiIT51Knh2OdPuAzD9o5s+fr9jYWB08eFB33HGHnnrqKT300EPq06ePvv/+e9ntdtlsNn333XdKSEiQdPEedQAAXA2bzaa5c+fq9ttvV+fOnfXee+95LJOk4OBglShRQi+88ILeeecdrVq1ioYbAHBVMs+Wymy4Fy5cqC+//FJLly7V0aNHVb9+fX3xxRfasWOH3nnnHY8j3oZhqGjRonI4HO45R3yVb53sXkDZbDbNnj1bPXv21EsvvaSKFStKkipVqqQhQ4ZIkh544AH997//1eHDhzV+/Hh3080pfQCAq7Vq1Sr17t1bPXr00IULFzRkyBDt3LlTo0aNUnBwsKSL12/Pnz9fy5cv18qVK9WgQQOTowYA+LPnn39ehQsX1tChQxUUFKTBgwfrq6++UnBwsIoUKSLDMDRr1iw1a9ZMX375pXr37q1Ro0bpqaeeUuvWrT36IF8/0k3T7QP279+vl19+WaNGjdKTTz7psaxixYoaOnSowsPD9e677yoqKkrLly93N+YAAFyJS2/Hcv78eQ0ZMkQvvPCCJOmuu+7S3XffLZfLpXfffVchISGKjo5Wo0aNNHnyZNWsWdPM0AEAfi4jI0OJiYnav3+/QkNDFRMTo99++03z5s1T1apVtW3bNr399ttq1aqVVq9erWbNmumrr75Shw4dVLVqVbVu3drsFPKEa7p9wJ9//qkOHTpo6tSpat68uSTPP4Yyr1k4evSoAgICVKJECTPDBQD4ucwas3btWu3evVtz5sxR7dq19fLLL7vX+eGHH3TXXXepX79+Gjt2rBwOh5KSklSqVCkTIwcA+LvMGpSenq6nnnpK27dvV7Vq1XTmzBlNmTLFvd6+ffv0+OOPS5K++eYbFStWTDt27ND111/v80e2/4lruk1y6b6OI0eO6NChQypWrJgkecwEu3HjRi1atEjp6emKjIyk4QYAXLXMa7ibNWumUaNG6ZtvvtHixYu1e/du9zqdOnXS7Nmz9cEHH+iZZ56RJBpuAMBVy+yDAgMD9f7776tmzZqaNWuWNm3a5F5mGIaio6PVtWtX7d69W+fOnZMkVa9e3S+u4f4nmu5rLHNDuvQahJYtW6pZs2Z6+OGHdeLECRUqVMi97LPPPtO8efPct3EBAOBKZdagQ4cO6YsvvtCnn36qRYsWacmSJVq3bp1eeeUV/fXXX+71O3bsqB9//NF9pAEAgKuVOWna119/raSkJI0dO1b333+/Tpw4oZEjRyo1NdXdK9WsWVNOp1MpKSker+FvR7o5vfwayjyV4ueff9b8+fN19uxZXXfddXr22We1YsUKDRs2TGlpafrggw908uRJrVixQp988ol++eUX1alTx+zwAQAW8PPPP2vChAlKSkrSRx99pEqVKkm6OJnajTfeqHvuuUdvvvmmypcvb3KkAAArMgxDJ06cUOnSpfXmm2/qmWee0fnz5zVw4ED9/vvvatWqlQYPHqzTp09r8ODBSktL04oVK7K9pZi/8N/I/ZDNZtOsWbPUpUsXnThxQpI0ZswYdenSRa1atdJbb72lChUqqFOnTnr66ae1fPlyLV++nIYbAJBvXC6XlixZohUrVmjz5s3u8ebNm2vFihWaM2eOnnjiCf39998mRgkAsLKSJUvq1Vdf1fTp0/Xnn38qKChI7733npo0aaLPPvtMTZs21YsvvqgSJUpo6dKlstvtfn3mL033NbR//369+OKLeuONN/TJJ58oNjZWFy5cUFRUlAICAtSmTRt9//33+vXXX/XTTz9p0aJFqlevntlhAwD8lGEY7uvejh8/rtTUVLVr104rV65UhQoV9Omnn+r33393r9+sWTMtWLBAa9eu5ZaUAIB8kdOJ1S1bttSpU6fc84kEBQVpzJgx6t+/vwzDULdu3TR16lQVKlRIGRkZfn2km9PLvezSWcj//PNPdenSRdu3b9dff/2lFi1aqHPnzpowYYIkacmSJerQoYOZ4QIALGDBggUqV66c6tatK0maNWuW3n77bR07dky1a9fWf/7zH9WoUUMdOnRQgwYNNHToUDVs2ND9/LS0NPf9uQEAyA/ff/+9IiIi3HdrkqSePXtq165dWrt2rbupPn/+vEaPHq3nn3/efYTbnxtuiSPdXmez2bR69WqNGzdOAQEBKlWqlObNm6dWrVqpc+fOev/99yVJ27Zt0xdffOFxxAEAgLw6cuSIBgwYoLFjx2rPnj3aunWr+vTpo9tvv12PPvqoypUrp65du+rnn3/W4sWLtWHDBr377rtas2aN+zWCgoJMzAAAYAWXng6+detWxcXFqW3btho8eLC++eYbSdKQIUMUHBysH3/8UZKUnp6uoKAgDRs2THa7XU6n0+8bbkkKMDsAq8vIyNBHH32kxMRE9erVy32qxP333+8+wi1dnKU8MTFRFSpUMDFaAIC/i4qK0syZM/XYY49p9OjRCg8P12OPPaYXX3xRkpSamqratWurf//++uGHHzR79my1adNGhQsXVr169RQUFMSp5QCAq+J0Ot0zjO/du1fR0dGaPXu2Nm/erHfffVfz58/XJ598on79+unw4cNaunSpbr31VgUGBnq8jr/NUp4TTi+/BrZv366GDRtq5syZqlSpkho3bqzOnTvrvvvuU2RkpGbMmKHJkyfr559/VkxMjNnhAgAsYMOGDfrPf/6jI0eOqEuXLu4zqyQpJSVFTz/9tNLS0jR16lStXLlSkZGRuv76602MGADg7z788EM1b97cPS/VsGHDNHv2bB0/fly9evXSwIEDVb58ee3du1eDBg1S4cKFNWPGDBUqVEhLly5Vq1atzE3AS2i689ml13BLcl+D8PTTT2vfvn367rvvtHTpUg0fPlyJiYkqUaKEwsPDNX78ePe1dwAA5IeEhATdcccdCg4O1tSpUz0m53zxxRc1b948rVmzhuu3AQBXLTExUW3atFGnTp303HPPaevWrXriiSf0/vvvKyEhQfPnz1fZsmUVGxur1q1bS5L++OMP/fLLLxo5cqQGDhyoF154wRLXcP8TTbcX/PTTT/rrr790//33uzeYWbNm6ZFHHtF3332ntm3bKikpSWfPnpXD4VCxYsUUGhpqctQAACv6448/9MADD6hhw4Z6+umn3Tt4H3vsMe3Zs0ezZ89WkSJFTI4SAGAFGzdu1COPPKLWrVvLbrerVq1a6tevnyRp3rx5Gj16tMLCwvT000+rbdu27ueNGzdOI0eO1Pbt21WyZEmzwvcamu58lp6erueff15jx47VnXfeqebNm2vIkCGSpEcffVSbN2/WokWLVKxYMZMjBQAUFPHx8erVq5fOnj2rNm3aKCgoSDNnztSSJUu4NSUAIF9t2LBBjz32mHbv3q3hw4fr6aefdi/LbLyLFy+uJ554Qu3bt5d08S5P99xzj6ZOnaqaNWuaFLn3WOu4vQ8IDAzU//73P23ZskVRUVH67LPPVLNmTU2aNEl16tRRRESENm7caHaYAIACpH79+poyZYrsdruWLl2qihUr6vfff6fhBgDkuwYNGmjixIkqXry4FixYoD/++MO9rEuXLhoyZIh27typJUuWuMenTJmizZs3q3jx4maE7HUc6faitLQ0nT59WkOHDtVff/2lLVu26ODBg3rqqac0duxYs8MDABQwv//+u4YNG6avv/5aERERZocDALCwTZs2qW/fvmrUqJEGDRqk2rVru5etXLlSTZs2dc9OPn36dF1//fVq2LChWeF6FU33NZKQkKBffvlFY8aM0cyZM5k0DQBgirS0NCZOAwBcE/Hx8XrkkUfc84rUqlXLY/mltxazMppuL/vnbObnz59XUFCQiREBAAAAwLURHx+vxx57TNHR0Xr77bdVqVIls0O65rim28subbglZbnhOwAAAABYVf369fX++++rWLFiio6ONjscU3CkGwAAAADgVZlnAFvxPtz/hqYbAAAAAOB1/7z0tqAoWLsYAAAAAACmKIgNt0TTDQAAAACA19B0AwAAAADgJTTdAAAAAAB4CU03AAAAAABeQtMNAAAAAICX0HQDAGARNptNs2fPNjsMAABwCZpuAAC8oE+fPrLZbLLZbCpUqJAqVaqk5557Tmlpabl+jRUrVshmsyk5OTlX6x86dEidOnW6woi9J/NzyOnnlVdeMTtEAAC8JsDsAAAAsKpbb71VkyZN0oULF/T777+rd+/estlseuutt/L1fdLT0xUYGKjSpUvn6+teTSyXOnTokPu/p0+fruHDh2vHjh3usaJFi16z+AAAuNY40g0AgJcEBQWpdOnSqlChgrp166YOHTpo8eLF7uUul0txcXGqVKmSQkJCVLduXc2cOVOStHfvXt14442SpOLFi8tms6lPnz6SpHbt2mnAgAF6+umnVapUKXXs2FFS1tPL//rrL913330KDw9XiRIldMcdd2jv3r2SpB9//FHBwcFZjqIPGjRIN910k/vxr7/+qtatWyskJEQVKlTQwIEDdebMGffyihUr6rXXXlOvXr0UGhqqRx99NMvnULp0afdPWFiYbDabSpcurWLFiqlatWpauHChx/qzZ89WkSJFdOrUKe3du1c2m03Tpk1TixYtFBwcrDp16uinn37yeM7mzZvVqVMnFS1aVFFRUXrooYeUlJSUi28JAADvoukGAOAa2Lx5s1auXOlxFDguLk5ffPGFJkyYoC1btmjw4MF68MEH9dNPP6lChQr69ttvJUk7duzQoUOHNHbsWPdzJ0+erMDAQP3222+aMGFClve7cOGCOnbsqGLFiumXX37Rb7/9pqJFi+rWW29Venq62rdvr/DwcPd7SJLT6dT06dP1wAMPSJJ2796tW2+9VXfffbcSEhI0ffp0/frrrxowYIDHe40aNUp169ZVfHy8Xn755Vx/JkWKFFGPHj00adIkj/FJkybpnnvuUbFixdxjzz77rJ555hnFx8erefPmuv3223X8+HFJUnJysm666SbVr19f69ev18KFC3XkyBHdd999uY4FAACvMQAAQL7r3bu34fh/7d1NSJTtHsfxn+PAkFIZ2JsOmiTVFGW+JLUoKUgCicJqYQZhE0koltGLBQUh0eQmKKRVmZhQUAxFLlxpWVJEVi7szWGiYiKC3Iw2Wfo/q4Yzx54nn2rycM73s5rhvl7+3Kv7d19zXZOYaMnJyeZyuUySORwOu3r1qpmZRSIRS0pKsu7u7ph+Xq/XysrKzMyso6PDJNnAwEBMm6KiIsvNzR0zpyTz+/1mZtbS0mLz58+30dHR6PXPnz/bpEmTrL293czM9uzZY2vWrIleb29vN5fLFZ3P6/Xarl27Yubo6uoyh8Nhnz59MjOzzMxM27hx47jvS1NTk02dOjX6/f79+5aYmGihUMjMzN6/f29Op9M6OzvNzCwYDJok8/l80T5fvnwxt9ttp06dMjOz+vp6Ky4ujpnnzZs3JsmeP38+7toAAIgH9nQDABAnq1ev1rlz5zQ4OKjTp0/L6XRq06ZNkqT+/n4NDQ1p7dq1MX2Gh4eVm5v7w7Hz8/P/9vqTJ0/U398fs1osSZFIRIFAQJJUXl6u5cuXKxQKKS0tTa2trSopKVFKSkp0jN7eXrW2tkb7m5lGR0cVDAbl8XgkSQUFBT+s968UFhZq0aJFam5uVl1dnS5duqTMzEytWrUqpt2KFSuin51OpwoKCvT06dNonR0dHd/dGx4IBDRv3ryfrg8AgF9F6AYAIE6Sk5OVnZ0tSbpw4YJycnJ0/vx5eb1ehcNhSVJbW5vS09Nj+rlcrnGN/XfC4bDy8/NjAvM306dPlyQtW7ZMc+fO1eXLl7V79275/X5dvHgxZozKykrV1NSMGSMjI2PctfzIzp071djYqLq6OjU1NamiokIJCQnj7h8Oh7V+/frvHlA3e/bsX6oNAIBfRegGAOAPcDgcOnLkiPbt26etW7dq4cKFcrlcev36tYqKir7b59v+75GRkX88X15enq5cuaIZM2ZoypQpf9muvLxcra2tcrvdcjgcKikpiRmjr68v+uIgXrZt26aDBw/qzJkz6uvr0/bt28e0uXfvXnT1++vXr3r48GF0b3leXp6uXbumOXPmyOnk0QYA8N+Fg9QAAPhDtmzZosTERDU2Nmry5Mnav3+/amtr1dzcrEAgoJ6eHp09e1bNzc2SpMzMTCUkJOjmzZv68OFDdHV8PMrLy5WamqoNGzaoq6tLwWBQnZ2dqqmp0du3b2Pa9fT06MSJE9q8eXPMKvuhQ4fU3d2t6upqPX78WC9fvtT169fHHKT2q6ZNm6bS0lIdOHBAxcXFcrvdY9o0NjbK7/fr2bNnqqqq0sDAgHbs2CFJqqqq0sePH1VWVqYHDx4oEAiovb1dFRUVP/XCAgCA34nQDQDAH+J0OlVdXa2GhgYNDg6qvr5eR48e1cmTJ+XxeLRu3Tq1tbUpKytLkpSenq7jx4+rrq5OM2fO/EdhNykpSbdv31ZGRoZKS0vl8Xjk9XoViURiVr6zs7NVWFio3t7e6Knl3yxZskS3bt3SixcvtHLlSuXm5urYsWNKS0v7PTfk33i9Xg0PD0eD9H/y+Xzy+XzKycnRnTt3dOPGDaWmpkqS0tLSdPfuXY2MjKi4uFiLFy/W3r17lZKSIoeDRx0AwMRKMDOb6CIAAMD/t5aWFtXW1ioUCsX8rdqrV6+UlZWlR48eaenSpRNXIAAAP4mNTwAAYMIMDQ3p3bt38vl8qqysjAncAAD8L+A3VwAAYMI0NDRowYIFmjVrlg4fPjzR5QAA8Nvx83IAAAAAAOKElW4AAAAAAOKE0A0AAAAAQJwQugEAAAAAiBNCNwAAAAAAcULoBgAAAAAgTgjdAAAAAADECaEbAAAAAIA4IXQDAAAAABAnhG4AAAAAAOLkX8fHocpULeDaAAAAAElFTkSuQmCC",
      "text/plain": [
       "<Figure size 1000x600 with 1 Axes>"
      ]
     },
     "metadata": {},
     "output_type": "display_data"
    }
   ],
   "source": [
    "# Run evaluation\n",
    "results = run_evaluation(n_samples=15, n_runs=3)\n",
    "\n",
    "# Print results\n",
    "for retriever_name, result in results.items():\n",
    "    if result and result.get(\"results\"):\n",
    "        print(f\"\\nResults for {retriever_name}:\")\n",
    "        df = result[\"results\"].to_pandas()\n",
    "        print(df)"
   ]
  }
 ],
 "metadata": {
  "kernelspec": {
   "display_name": "rag6998",
   "language": "python",
   "name": "python3"
  },
  "language_info": {
   "codemirror_mode": {
    "name": "ipython",
    "version": 3
   },
   "file_extension": ".py",
   "mimetype": "text/x-python",
   "name": "python",
   "nbconvert_exporter": "python",
   "pygments_lexer": "ipython3",
   "version": "3.11.10"
  }
 },
 "nbformat": 4,
 "nbformat_minor": 2
}