
RLHF-Reward-Modeling
Recipes to train reward model for RLHF.
Stars: 568

This repository, RLHF-Reward-Modeling, is dedicated to training reward models for DRL-based RLHF (PPO), Iterative SFT, and iterative DPO. It provides state-of-the-art performance in reward models with a base model size of up to 13B. The installation instructions involve setting up the environment and aligning the handbook. Dataset preparation requires preprocessing conversations into a standard format. The code can be run with Gemma-2b-it, and evaluation results can be obtained using provided datasets. The to-do list includes various reward models like Bradley-Terry, preference model, regression-based reward model, and multi-objective reward model. The repository is part of iterative rejection sampling fine-tuning and iterative DPO.
README:
Our models and codes have contributed to many academic research projects, e.g.,
- Xu Zhangchen, et al. "Magpie: Alignment Data Synthesis from Scratch by Prompting Aligned LLMs with Nothing."
- Chen, Lichang, et al. "OPTune: Efficient Online Preference Tuning."
- Xie, Tengyang, et al. "Exploratory Preference Optimization: Harnessing Implicit Q*-Approximation for Sample-Efficient RLHF." arXiv preprint arXiv:2405.21046 (2024).
- Zhong, Han, et al. "Dpo meets ppo: Reinforced token optimization for rlhf." arXiv preprint arXiv:2404.18922 (2024).
- Zheng, Chujie, et al. "Weak-to-strong extrapolation expedites alignment." arXiv preprint arXiv:2404.16792 (2024).
- Ye, Chenlu, et al. "A theoretical analysis of nash learning from human feedback under general kl-regularized preference." arXiv preprint arXiv:2402.07314 (2024).
- Chen, Ruijun, et al. "Self-Evolution Fine-Tuning for Policy Optimization"
- Li Bolian, et al., Cascade Reward Sampling for Efficient Decoding-Time Alignment
- Zhang, Yuheng, et al. "Iterative Nash Policy Optimization: Aligning LLMs with General Preferences via No-Regret Learning"
- Lin Tzu-Han, et al., "DogeRM: Equipping Reward Models with Domain Knowledge through Model Merging",
- Yang Rui, et al., "Regularizing Hidden States Enables Learning Generalizable Reward Model for LLMs"
- Junsoo Park, et al., "OffsetBias: Leveraging Debiased Data for Tuning Evaluators"
- Meng Yu, et al., "SimPO: Simple Preference Optimization with a Reference-Free Reward"
- Song Yifan, et al., "The Good, The Bad, and The Greedy: Evaluation of LLMs Should Not Ignore Non-Determinism"
π₯π₯π₯π₯π₯π₯π₯π₯π₯π₯π₯π₯π₯π₯π₯π₯π₯π₯π₯π₯π₯π₯
π Our ArmoRM is the Rank #1 8B model on RewardBench!
π The top-3 open-source 8B reward models on RewardBench (ArmoRM, Pair Pref. Model, BT RM) are all trained with this repo!
π The pairwise preference model training code is available (pair-pm/
)!
π₯π₯π₯π₯π₯π₯π₯π₯π₯π₯π₯π₯π₯π₯π₯π₯π₯π₯π₯π₯π₯π₯
-
Tech Report
-
Models:
- Absolute-Rating Multi-Objective Reward Model (ArmoRM): ArmoRM-Llama3-8B-v0.1
- Pairwise Preference Reward Model: pair-preference-model-LLaMA3-8B
- Bradley-Terry Reward Model: FsfairX-LLaMA3-RM-v0.1
-
Architectures
-
Model Base Model Method Score Chat Chat Hard Safety Reasoning Prior Sets (0.5 weight) ArmoRM-Llama3-8B-v0.1 (Ours) Llama-3 8B ArmoRM + MoE 89.0 96.9 76.8 92.2 97.3 74.3 Cohere May 2024 Unknown Unknown 88.2 96.4 71.3 92.7 97.7 78.2 pair-preference-model (Ours) Llama-3 8B SliC-HF 85.7 98.3 65.8 89.7 94.7 74.6 GPT-4 Turbo (0125 version) GPT-4 Turbo LLM-as-a-Judge 84.3 95.3 74.3 87.2 86.9 70.9 FsfairX-LLaMA3-RM-v0.1 (Ours) Llama-3 8B Bradley-Terry 83.6 99.4 65.1 87.8 86.4 74.9 Starling-RM-34B Yi-34B Bradley-Terry 81.4 96.9 57.2 88.2 88.5 71.4 -
Evaluation Results (from RLHF Workflow)
TL;DL: this is a repo for training the reward/preference model for DRL-based RLHF (PPO), Iterative SFT (Rejection sampling fine-tuning), and iterative DPO.
- 4 x A40 48G: we can train Gemma-7B-it with max_length 4096 by Deepspeed Zero-3 + gradient checkpoint;
- 4 x A100 80G: we can train Gemma-7B-it with max_length 4096 by gradient checkpoint;
- The resulting reward models achieve SOTA performance as open-source RMs in the leaderboard of RewardBench.
- Check out our blog post!
It is recommeded to create separate environmnets for the Bradley-Terry reward model and pair wise preference model. The installation instructions are provided in the corresponding folders.
The dataset should be preprocessed as the standard format, where each of the sample consists of two conversations 'chosen' and 'rejected' and they share the same prompt. Here is an example of the rejected sample in the comparison pair.
[
{ "content": "Please identify the top 5 rarest animals in the world.", "role": "user" },
{ "content": "Do you mean animals that are really rare, or rare relative to the size of the human population?", "role": "assistant" },
{ "content": "The ones that are really rare.", "role": "user" },
{ "content": "Alright, hereβs what I found:", "role": "assistant" },
]
We preprocess many open-source preference datasets into the standard format and upload them to the hugginface hub. You can find them HERE. We have also searched and founda that some of the following mixture of preference dataset useful.
- weqweasdas/preference_dataset_mix2
- weqweasdas/preference_dataset_mixture2_and_safe_pku
- hendrydong/preference_700K where the details can be found in the dataset card.
You can evaluate the resulting reward model with the dataset provided by benchmark by the following command.
CUDA_VISIBLE_DEVICES=1 python ./useful_code/eval_reward_bench_bt.py --reward_name_or_path ./models/gemma_2b_mixture2_last_checkpoint --record_dir ./bench_mark_eval.txt
- [x] Bradley-Terry Reward Model
- [x] Preference model
- [x] Multi-Objective Reward Model
- [ ] LLM-as-a-judge
The repo was part of the iterative rejection sampling fine-tuning and iterative DPO. If you find the content of this repo useful in your work, please consider citing:
@article{dong2024rlhf,
title={RLHF Workflow: From Reward Modeling to Online RLHF},
author={Dong, Hanze and Xiong, Wei and Pang, Bo and Wang, Haoxiang and Zhao, Han and Zhou, Yingbo and Jiang, Nan and Sahoo, Doyen and Xiong, Caiming and Zhang, Tong},
journal={arXiv preprint arXiv:2405.07863},
year={2024}
}
@article{ArmoRM,
title={Interpretable Preferences via Multi-Objective Reward Modeling and Mixture-of-Experts},
author={Haoxiang Wang and Wei Xiong and Tengyang Xie and Han Zhao and Tong Zhang},
journal={arXiv preprint arXiv:2406.12845},
}
@article{dong2023raft,
title={{RAFT}: Reward rAnked FineTuning for Generative Foundation Model Alignment},
author={Hanze Dong and Wei Xiong and Deepanshu Goyal and Yihan Zhang and Winnie Chow and Rui Pan and Shizhe Diao and Jipeng Zhang and KaShun SHUM and Tong Zhang},
journal={Transactions on Machine Learning Research},
issn={2835-8856},
year={2023},
url={https://openreview.net/forum?id=m7p5O7zblY},
}
@article{xiong2024iterative,
title={Iterative Preference Learning from Human Feedback: Bridging Theory and Practice for RLHF under KL-Constraint},
author={Wei Xiong and Hanze Dong and Chenlu Ye and Ziqi Wang and Han Zhong and Heng Ji and Nan Jiang and Tong Zhang},
year={2024},
journal={ICML}
}
For Tasks:
Click tags to check more tools for each tasksFor Jobs:
Alternative AI tools for RLHF-Reward-Modeling
Similar Open Source Tools

RLHF-Reward-Modeling
This repository, RLHF-Reward-Modeling, is dedicated to training reward models for DRL-based RLHF (PPO), Iterative SFT, and iterative DPO. It provides state-of-the-art performance in reward models with a base model size of up to 13B. The installation instructions involve setting up the environment and aligning the handbook. Dataset preparation requires preprocessing conversations into a standard format. The code can be run with Gemma-2b-it, and evaluation results can be obtained using provided datasets. The to-do list includes various reward models like Bradley-Terry, preference model, regression-based reward model, and multi-objective reward model. The repository is part of iterative rejection sampling fine-tuning and iterative DPO.

rageval
Rageval is an evaluation tool for Retrieval-augmented Generation (RAG) methods. It helps evaluate RAG systems by performing tasks such as query rewriting, document ranking, information compression, evidence verification, answer generation, and result validation. The tool provides metrics for answer correctness and answer groundedness, along with benchmark results for ASQA and ALCE datasets. Users can install and use Rageval to assess the performance of RAG models in question-answering tasks.

VILA
VILA is a family of open Vision Language Models optimized for efficient video understanding and multi-image understanding. It includes models like NVILA, LongVILA, VILA-M3, VILA-U, and VILA-1.5, each offering specific features and capabilities. The project focuses on efficiency, accuracy, and performance in various tasks related to video, image, and language understanding and generation. VILA models are designed to be deployable on diverse NVIDIA GPUs and support long-context video understanding, medical applications, and multi-modal design.

MiniCPM-V
MiniCPM-V is a series of end-side multimodal LLMs designed for vision-language understanding. The models take image and text inputs to provide high-quality text outputs. The series includes models like MiniCPM-Llama3-V 2.5 with 8B parameters surpassing proprietary models, and MiniCPM-V 2.0, a lighter model with 2B parameters. The models support over 30 languages, efficient deployment on end-side devices, and have strong OCR capabilities. They achieve state-of-the-art performance on various benchmarks and prevent hallucinations in text generation. The models can process high-resolution images efficiently and support multilingual capabilities.

FlagEmbedding
FlagEmbedding focuses on retrieval-augmented LLMs, consisting of the following projects currently: * **Long-Context LLM** : Activation Beacon * **Fine-tuning of LM** : LM-Cocktail * **Embedding Model** : Visualized-BGE, BGE-M3, LLM Embedder, BGE Embedding * **Reranker Model** : llm rerankers, BGE Reranker * **Benchmark** : C-MTEB

inference
Xorbits Inference (Xinference) is a powerful and versatile library designed to serve language, speech recognition, and multimodal models. With Xorbits Inference, you can effortlessly deploy and serve your or state-of-the-art built-in models using just a single command. Whether you are a researcher, developer, or data scientist, Xorbits Inference empowers you to unleash the full potential of cutting-edge AI models.

SynapseML
SynapseML (previously known as MMLSpark) is an open-source library that simplifies the creation of massively scalable machine learning (ML) pipelines. It provides simple, composable, and distributed APIs for various machine learning tasks such as text analytics, vision, anomaly detection, and more. Built on Apache Spark, SynapseML allows seamless integration of models into existing workflows. It supports training and evaluation on single-node, multi-node, and resizable clusters, enabling scalability without resource wastage. Compatible with Python, R, Scala, Java, and .NET, SynapseML abstracts over different data sources for easy experimentation. Requires Scala 2.12, Spark 3.4+, and Python 3.8+.

openrl
OpenRL is an open-source general reinforcement learning research framework that supports training for various tasks such as single-agent, multi-agent, offline RL, self-play, and natural language. Developed based on PyTorch, the goal of OpenRL is to provide a simple-to-use, flexible, efficient and sustainable platform for the reinforcement learning research community. It supports a universal interface for all tasks/environments, single-agent and multi-agent tasks, offline RL training with expert dataset, self-play training, reinforcement learning training for natural language tasks, DeepSpeed, Arena for evaluation, importing models and datasets from Hugging Face, user-defined environments, models, and datasets, gymnasium environments, callbacks, visualization tools, unit testing, and code coverage testing. It also supports various algorithms like PPO, DQN, SAC, and environments like Gymnasium, MuJoCo, Atari, and more.

AIOS
AIOS, a Large Language Model (LLM) Agent operating system, embeds large language model into Operating Systems (OS) as the brain of the OS, enabling an operating system "with soul" -- an important step towards AGI. AIOS is designed to optimize resource allocation, facilitate context switch across agents, enable concurrent execution of agents, provide tool service for agents, maintain access control for agents, and provide a rich set of toolkits for LLM Agent developers.

pytorch-grad-cam
This repository provides advanced AI explainability for PyTorch, offering state-of-the-art methods for Explainable AI in computer vision. It includes a comprehensive collection of Pixel Attribution methods for various tasks like Classification, Object Detection, Semantic Segmentation, and more. The package supports high performance with full batch image support and includes metrics for evaluating and tuning explanations. Users can visualize and interpret model predictions, making it suitable for both production and model development scenarios.

AceCoder
AceCoder is a tool that introduces a fully automated pipeline for synthesizing large-scale reliable tests used for reward model training and reinforcement learning in the coding scenario. It curates datasets, trains reward models, and performs RL training to improve coding abilities of language models. The tool aims to unlock the potential of RL training for code generation models and push the boundaries of LLM's coding abilities.

RAG-Retrieval
RAG-Retrieval provides full-chain RAG retrieval fine-tuning and inference code. It supports fine-tuning any open-source RAG retrieval models, including vector (embedding, graph a), delayed interactive models (ColBERT, graph d), interactive models (cross encoder, graph c). For inference, RAG-Retrieval focuses on ranking (reranker) and has developed a lightweight Python library rag-retrieval, providing a unified way to call any different RAG ranking models.

UHGEval
UHGEval is a comprehensive framework designed for evaluating the hallucination phenomena. It includes UHGEval, a framework for evaluating hallucination, XinhuaHallucinations dataset, and UHGEval-dataset pipeline for creating XinhuaHallucinations. The framework offers flexibility and extensibility for evaluating common hallucination tasks, supporting various models and datasets. Researchers can use the open-source pipeline to create customized datasets. Supported tasks include QA, dialogue, summarization, and multi-choice tasks.

UMbreLLa
UMbreLLa is a tool designed for deploying Large Language Models (LLMs) for personal agents. It combines offloading, speculative decoding, and quantization to optimize single-user LLM deployment scenarios. With UMbreLLa, 70B-level models can achieve performance comparable to human reading speed on an RTX 4070Ti, delivering exceptional efficiency and responsiveness, especially for coding tasks. The tool supports deploying models on various GPUs and offers features like code completion and CLI/Gradio chatbots. Users can configure the LLM engine for optimal performance based on their hardware setup.

Cherry_LLM
Cherry Data Selection project introduces a self-guided methodology for LLMs to autonomously discern and select cherry samples from open-source datasets, minimizing manual curation and cost for instruction tuning. The project focuses on selecting impactful training samples ('cherry data') to enhance LLM instruction tuning by estimating instruction-following difficulty. The method involves phases like 'Learning from Brief Experience', 'Evaluating Based on Experience', and 'Retraining from Self-Guided Experience' to improve LLM performance.

Pai-Megatron-Patch
Pai-Megatron-Patch is a deep learning training toolkit built for developers to train and predict LLMs & VLMs by using Megatron framework easily. With the continuous development of LLMs, the model structure and scale are rapidly evolving. Although these models can be conveniently manufactured using Transformers or DeepSpeed training framework, the training efficiency is comparably low. This phenomenon becomes even severer when the model scale exceeds 10 billion. The primary objective of Pai-Megatron-Patch is to effectively utilize the computational power of GPUs for LLM. This tool allows convenient training of commonly used LLM with all the accelerating techniques provided by Megatron-LM.
For similar tasks

RLHF-Reward-Modeling
This repository, RLHF-Reward-Modeling, is dedicated to training reward models for DRL-based RLHF (PPO), Iterative SFT, and iterative DPO. It provides state-of-the-art performance in reward models with a base model size of up to 13B. The installation instructions involve setting up the environment and aligning the handbook. Dataset preparation requires preprocessing conversations into a standard format. The code can be run with Gemma-2b-it, and evaluation results can be obtained using provided datasets. The to-do list includes various reward models like Bradley-Terry, preference model, regression-based reward model, and multi-objective reward model. The repository is part of iterative rejection sampling fine-tuning and iterative DPO.

MathEval
MathEval is a benchmark designed for evaluating the mathematical capabilities of large models. It includes over 20 evaluation datasets covering various mathematical domains with more than 30,000 math problems. The goal is to assess the performance of large models across different difficulty levels and mathematical subfields. MathEval serves as a reliable reference for comparing mathematical abilities among large models and offers guidance on enhancing their mathematical capabilities in the future.

AIGC_text_detector
AIGC_text_detector is a repository containing the official codes for the paper 'Multiscale Positive-Unlabeled Detection of AI-Generated Texts'. It includes detector models for both English and Chinese texts, along with stronger detectors developed with enhanced training strategies. The repository provides links to download the detector models, datasets, and necessary preprocessing tools. Users can train RoBERTa and BERT models on the HC3-English dataset using the provided scripts.

RLHF-Reward-Modeling
This repository contains code for training reward models for Deep Reinforcement Learning-based Reward-modulated Hierarchical Fine-tuning (DRL-based RLHF), Iterative Selection Fine-tuning (Rejection sampling fine-tuning), and iterative Decision Policy Optimization (DPO). The reward models are trained using a Bradley-Terry model based on the Gemma and Mistral language models. The resulting reward models achieve state-of-the-art performance on the RewardBench leaderboard for reward models with base models of up to 13B parameters.

AceCoder
AceCoder is a tool that introduces a fully automated pipeline for synthesizing large-scale reliable tests used for reward model training and reinforcement learning in the coding scenario. It curates datasets, trains reward models, and performs RL training to improve coding abilities of language models. The tool aims to unlock the potential of RL training for code generation models and push the boundaries of LLM's coding abilities.
For similar jobs

sweep
Sweep is an AI junior developer that turns bugs and feature requests into code changes. It automatically handles developer experience improvements like adding type hints and improving test coverage.

teams-ai
The Teams AI Library is a software development kit (SDK) that helps developers create bots that can interact with Teams and Microsoft 365 applications. It is built on top of the Bot Framework SDK and simplifies the process of developing bots that interact with Teams' artificial intelligence capabilities. The SDK is available for JavaScript/TypeScript, .NET, and Python.

ai-guide
This guide is dedicated to Large Language Models (LLMs) that you can run on your home computer. It assumes your PC is a lower-end, non-gaming setup.

classifai
Supercharge WordPress Content Workflows and Engagement with Artificial Intelligence. Tap into leading cloud-based services like OpenAI, Microsoft Azure AI, Google Gemini and IBM Watson to augment your WordPress-powered websites. Publish content faster while improving SEO performance and increasing audience engagement. ClassifAI integrates Artificial Intelligence and Machine Learning technologies to lighten your workload and eliminate tedious tasks, giving you more time to create original content that matters.

chatbot-ui
Chatbot UI is an open-source AI chat app that allows users to create and deploy their own AI chatbots. It is easy to use and can be customized to fit any need. Chatbot UI is perfect for businesses, developers, and anyone who wants to create a chatbot.

BricksLLM
BricksLLM is a cloud native AI gateway written in Go. Currently, it provides native support for OpenAI, Anthropic, Azure OpenAI and vLLM. BricksLLM aims to provide enterprise level infrastructure that can power any LLM production use cases. Here are some use cases for BricksLLM: * Set LLM usage limits for users on different pricing tiers * Track LLM usage on a per user and per organization basis * Block or redact requests containing PIIs * Improve LLM reliability with failovers, retries and caching * Distribute API keys with rate limits and cost limits for internal development/production use cases * Distribute API keys with rate limits and cost limits for students

uAgents
uAgents is a Python library developed by Fetch.ai that allows for the creation of autonomous AI agents. These agents can perform various tasks on a schedule or take action on various events. uAgents are easy to create and manage, and they are connected to a fast-growing network of other uAgents. They are also secure, with cryptographically secured messages and wallets.

griptape
Griptape is a modular Python framework for building AI-powered applications that securely connect to your enterprise data and APIs. It offers developers the ability to maintain control and flexibility at every step. Griptape's core components include Structures (Agents, Pipelines, and Workflows), Tasks, Tools, Memory (Conversation Memory, Task Memory, and Meta Memory), Drivers (Prompt and Embedding Drivers, Vector Store Drivers, Image Generation Drivers, Image Query Drivers, SQL Drivers, Web Scraper Drivers, and Conversation Memory Drivers), Engines (Query Engines, Extraction Engines, Summary Engines, Image Generation Engines, and Image Query Engines), and additional components (Rulesets, Loaders, Artifacts, Chunkers, and Tokenizers). Griptape enables developers to create AI-powered applications with ease and efficiency.