Corrected Context Template and Instruct Mode settings for SillyTavern.
Pandora from MistralAI has opened a PR for SillyTavern to add corrected templates that properly accommodate all of Mistral's models (as their tokenizer behavior is slightly different) but for Nemo in particular, these should be functional.
Pandora's templates are going to look different to the ones in this repo. If you want to use them, the equivalent files in Pandora's repo are here (Context) and here (Instruct).
Until the PR is merged, though, they're likely to change.
The new Context Template and Instruct Mode presets for all Mistral architectures have been merged to SillyTavern's Staging branch. The current versions of the templates are now hosted in this repository, for if you don't want to/can't update. The new update to SillyTavern Staging introduced 3 new Mistral entries to both Context Template and Instruct Mode.
Use the following SillyTavern templates for the respective Mistral model architectures:
Mistral V1:
- Mistral 7B v1
- Mistral 7B v2
- Mixtral 8x7B
Mistral V2 & V3:
- Mistral 7B v3
- Codestral
- Mixtral 8x22B
- Mistral Small
- Mistral Medium (incl. Miqu, probably)
- Mistral Large
- Pixtral
Mistral V3-Tekken:
- Mistral Nemo
If you're using something based on Mistral 7B and you can't tell which version it is, then just use whichever one works better, I guess. Tell the maker of your model to add the base_model parameter in the YAML section of their model card, so that HuggingFace's model tree will work properly.