3 ms·
> The most promising idea is to use reasoning models to generate data, and then train our non-reasoning models with the reasoning-embedded data. DeepSeek did p
by sheepscreek 2y ago
> The most promising idea is to use reasoning models to generate data, and then train our non-reasoning models with the reasoning-embedded data.
DeepSeek did precisely this with their LLama fine-tunes. You can try the 70B one here (might have to sign up):
https://groq.com/groqcloud-makes-deepseek-r1-distill-llama-70b-available/ https://groq.com/groqcloud-makes-deepseek-r1-distill-llama-7...
- spyckie2 2y agoYes, but I meant it slightly differently than the distills. The idea is to create the next gen SOTA non reasoning model with synthetic reasoning training data.
- entropicdrifter 2y agoSo you mean something like, "what if the baseline, off-the-cuff response for the next-gen models was tuned based on the results of the reasoning model excluding the reasoning itself?"
- spyckie2 2y agoExactly, albeit it may need the reasoning later to form the proper foundational logic in the weights.