RLHFlow/RLHF-Reward-Modeling
A collection of training recipes for reward models used in RLHF-based LLM alignment, including Bradley-Terry, pairwise, multi-objective, and process-supervised approaches.
Not currently ranked — collecting fresh signals.
star history
The repository provides implementations of various reward modeling techniques for training LLMs via Reinforcement Learning from Human Feedback (RLHF). It includes classic Bradley-Terry reward modeling, pairwise preference models that predict response preference from prompt-response pairs, multi-objective reward models using mixture-of-experts aggregation, and process-supervised reward models for mathematical reasoning. Each approach includes code, data, hyperparameters, and references to associated research papers.
Frequently asked
- What is RLHFlow/RLHF-Reward-Modeling?
- A collection of training recipes for reward models used in RLHF-based LLM alignment, including Bradley-Terry, pairwise, multi-objective, and process-supervised approaches.
- Is RLHF-Reward-Modeling open source?
- Yes — RLHFlow/RLHF-Reward-Modeling is open source, released under the Apache-2.0 license.
- What language is RLHF-Reward-Modeling written in?
- RLHFlow/RLHF-Reward-Modeling is primarily written in Python.
- How popular is RLHF-Reward-Modeling?
- RLHFlow/RLHF-Reward-Modeling has 1.5k stars on GitHub.
- Where can I find RLHF-Reward-Modeling?
- RLHFlow/RLHF-Reward-Modeling is on GitHub at https://github.com/RLHFlow/RLHF-Reward-Modeling.