3 ms·
My bet is that they use the pre-assembled training data to train the reward LLM and maybe also for fine tuning the model in addition to RLHF. This is my best g
by whimsicalism 4y ago
My bet is that they use the pre-assembled training data to train the reward LLM and maybe also for fine tuning the model in addition to RLHF.
This is my best guess, afaik the research is pretty opaque right now.