PPO-M (PPO with Calibrated Reward Modeling) is an RLHF algorithm to mitigate verbalized overconfidence in RLHF-trained Large Language Models. PPO-M calibrates the reward modeling process by augmenting the binary pairwise ranking dataset with explicit confidence scores, and encourages the reward model to align confidence levels with response quality. Please refer to our preprint (Taming Overconfidence in LLMs: Reward Calibration in RLHF) and repo for more details.

We train teknium/OpenHermes-2.5-Mistral-7B on our HINT-lab/prompt-collections-final-v0.3 with our calibrated reward model HINT-lab/mistral-7b-hermes-crm-skywork.