3 ms·
Interested to see what folks do with putting DeepSeek-style RL methods on top of this. The smaller Mistral models have always punched above their weight and bee
by bugglebeetle 2y ago
Interested to see what folks do with putting DeepSeek-style RL methods on top of this. The smaller Mistral models have always punched above their weight and been the best for fine-tuning.
- petercooper 2y agoIt's not RL, but you can get a long way with a thorough system prompt to encourage it to engage in 'thinking' behavior on its own without extra training. Just playing with it myself now with promising results - Mistral Small seems very receptive to this approach (not all models are - cough, Llama). Update: This is such a prompt: https://gist.github.com/peterc/955d797ee35b3c777d76a2d881d2fb63 https://gist.github.com/peterc/955d797ee35b3c777d76a2d881d2f...