# RLHFlow/RLHF-Reward-Modeling

Recipes to train reward model for RLHF.

Repository: https://github.com/RLHFlow/RLHF-Reward-Modeling
Canonical: https://ross.abutalabs.com/products/rlhf-reward-modeling
Homepage: https://rlhflow.github.io/
Language: Python
License: Apache-2.0
License Family: permissive
Topics: llm, rlhf, reward-models, llama3
Last push: 2025-04-24T11:39:23+00:00

## Health v2 (maintenance only)
Score: 33/100 (v2, computed 2026-09-02T17:46:02.011165+00:00)
- activity 18, release rhythm 35, longevity 63
- inputs: {"age_days": 895, "days_push": 496, "days_rel": null, "gap_med": null, "n_releases_24m": 0}
- flags: no_releases
- formula: round(0.45*activity + 0.35*rhythm + 0.20*longevity); archived -> min(score, 10)

## Adoption (not part of the score)
Stars 1541, forks 110 (observed 2026-08-28T04:05:00.663530+00:00)

## What it is
A collection of training recipes for reward models used in RLHF, covering Bradley-Terry reward models, pairwise preference models, ArmoRM, process/outcome-supervised math reward models, and decision-tree reward models. It provides reproducible code, data, and hyperparameters for state-of-the-art reward modeling research.

## Use cases
- train a bradley-terry reward model for rlhf
- train a pairwise preference model on human preference data
- build a multi-objective reward model with mixture-of-experts aggregation
- train process-supervised and outcome-supervised reward models for math
- reduce length bias in reward modeling
- reproduce state-of-the-art rewardbench results

## When to choose
- you need to train a reward model for RLHF fine-tuning of LLMs
- you want reproducible recipes for multiple reward modeling techniques
- you are doing research on preference modeling or reward hacking mitigation

## When to avoid
- you just want to run inference with an existing reward model
- you need a full end-to-end RLHF training pipeline including PPO or DPO
- you are not working with large language models

## Facets
- artifact type: library
- maturity: active
- function: machine-learning, llm-training, deep-learning
- domain: large-language-models, reinforcement-learning, machine-learning, artificial-intelligence
- platform: python
- tags: rlhf, reward-model, bradley-terry, preference-model, prm, armorm, rewardbench, llama3, gpu, linux

## Member repositories
- RLHFlow/RLHF-Reward-Modeling (main) score 33

## Provenance
- Observed fields: from GitHub, fetched 2026-08-28T04:05:00.663530+00:00.
- Health v2: computed from the inputs above; adoption is never an input.
- Inferred fields (summary, facets, guidance): AI-extracted, prompt v1, taxonomy v1, on 2026-08-30T04:30:47.081200+00:00, confidence not recorded.
  - readme: https://github.com/RLHFlow/RLHF-Reward-Modeling (fetched 2026-08-28T04:05:00.663530+00:00, sha 31e127a042a9)
  - homepage: https://rlhflow.github.io/ (fetched 2026-08-29T11:32:14.777976+00:00, sha 7eeeb8336f4a)
- Data as of 2026-08-30T08:39:29.467469+00:00.