Closed natolambert closed 1 month ago
Would be interesting to test after IFT and before PPO particularly. Thinking about it, rejection sampling helps the model prioritize the distribution that the reward model likes, which sounds similar to share prompts on the RM and PPO stages.
Would be interesting to test after IFT and before PPO particularly. Thinking about it, rejection sampling helps the model prioritize the distribution that the reward model likes, which sounds similar to share prompts on the RM and PPO stages.