We train a collection of models under RLHF on the above datasets. We use DPO for hh-rlhf and unalignment, and train a PPO on completing IMDB prefixes with positive sentiment.
We train a collection of models under RLHF on the above datasets. We use DPO for hh-rlhf and unalignment, and train a PPO on completing IMDB prefixes with positive sentiment.