RLHF-Reward-Modeling

Why we included this project

Reward model training is often the fiddliest part of an RLHF pipeline, and this repository collects the recipes the RLHFlow team used to train their released models. It covers the classic Bradley-Terry reward model and pairwise preference models, then goes further with research variants like ArmoRM's multi-objective mixture-of-experts aggregation, length-bias-corrected reward modeling, and semi-supervised self-training. The training data, code, and hyperparameters for each approach are included, so you can reproduce the results on your own preference data instead of reconstructing the setup from a paper. The decision-tree reward model is a nice option if you want a scoring signal you can actually inspect rather than a black-box network. For teams already at home with Hugging Face training loops, this is a useful reference for comparing reward model designs side by side.