TheDrummer's picture
Update README.md
b958941 verified
![image/png](https://cdn-uploads.huggingface.co/production/uploads/65f2fd1c25b848bd061b5c2e/il5PmyJOCwkDR_1dwfzHa.png)
Mistral (Non-Tekken), i.e., Mistral v3 + `[SYSTEM_PROMPT]`
Looks like `<think>` doesn't need to be prefilled. You can opt out of reasoning - it's just as good, maybe even better.
No toxic data so you may want to prefill / guide reasoning when dealing with heavy themes (unless your prompt is sufficiently instructed/gaslit to be evil).
Alternatively, since `<think>` is not a special token, you can influence reasoning by phrasing it like `<evil_think>`, `<creative_think>`, or `<spicy_think>`, etc. It's smart enough to close it properly.
Yes, this is how much I want to avoid tuning MoEs.
![image/png](https://cdn-uploads.huggingface.co/production/uploads/65f2fd1c25b848bd061b5c2e/Rson4ntOeqeYOTeKodMxH.png)