Online-RLHF

A recipe for online RLHF and online iterative DPO.

Stars: 386

Visit

This repository, Online RLHF, focuses on aligning large language models (LLMs) through online iterative Reinforcement Learning from Human Feedback (RLHF). It aims to bridge the gap in existing open-source RLHF projects by providing a detailed recipe for online iterative RLHF. The workflow presented here has shown to outperform offline counterparts in recent LLM literature, achieving comparable or better results than LLaMA3-8B-instruct using only open-source data. The repository includes model releases for SFT, Reward model, and RLHF model, along with installation instructions for both inference and training environments. Users can follow step-by-step guidance for supervised fine-tuning, reward modeling, data generation, data annotation, and training, ultimately enabling iterative training to run automatically.

README:

Online RLHF

TL;DL: this is a repo to align the large language models (LLMs) by online iterative RLHF. Also check out our technical report and Huggingface Repo!

We present the workflow of Online Iterative Reinforcement Learning from Human Feedback (RLHF), which is widely reported to outperform its offline counterpart by a large margin in the recent LLM literature. However, existing open-source RLHF projects are still largely confined to the offline learning setting. In this repo, we aim to fill in this gap and provide a detailed recipe that is easy to be reproduced for online iterative RLHF. In particular, with our recipe, with only open-source data, we can achieve comparable or even better results than LLaMA3-8B-instruct.

Model Releases

SFT model, also check more SFT checkpoints Here
Reward model, also check more reward models Here
RLHF model

Installation instructions

It is recommended to have two separate environments for inference and training, respectively.

Note that the numpy version should be numpy<2.0. Numpy 2.0 will encounter unexpected issues!!!

Inference Environment

conda create -n vllm python=3.10.9
conda activate vllm
pip install datasets

# The following code is tested for CUDA12.0-12.2, and CUDA12.6
# To develop llama-3, mistral, gemma-1, 1.1, 2, deepseek you can consider the following vllm version
pip install vllm==0.5.4

pip install accelerate==0.33.0
pip install deepspeed==0.14.5
pip install transformers==4.43.4
pip install numpy==1.26.4 #Note that the numpy version should be `numpy<2.0`.  `Numpy 2.0` will encounter unexpected issues!!!

Training Environment

conda create -n rlhflow python=3.10.9
conda activate rlhflow

git clone https://github.com/huggingface/alignment-handbook.git
cd ./alignment-handbook/
git checkout 27f7dbf00663dab66ad7334afb7a1311fa251f41
pip3 install torch==2.1.2 torchvision torchaudio
python -m pip install .
pip install flash-attn==2.6.3
pip install accelerate==0.33.0
pip install huggingface-hub==0.24.7

You also need to install the wandb to record the training and login with your huggingface account so that you have access to the LLaMA3 models.

pip install wandb==0.17.7

wandb login
huggingface-cli login

Get Started

We present a step-by-step guidance in this section.

Step 1 Supervised Fine-tuning

To start with, you should first preprocess your dataset into the standard format. Here is an example of the dataset. You may need to adjust the hyper-parameters (batch size, packing size) according to your computational resources. To run SFT, you can use the following command.

# You can adjust the training parameters in ./sft/sft.py
accelerate launch ./sft/sft.py

# Train with deepspeed stage3 
# You may need to adjust ./configs/zero3.yaml, especially the num_processes (the number of GPUs) according to your environment
accelerate launch --config_file ./configs/zero3.yaml ./sft/sft.py

Step 2 Reward Modeling

We refer the interested readers to this repo for a detailed recipe to train the state-of-the-art open-source reward/preference models. We have trained several RMs and prepared them on the huggingface like sfairXC/FsfairX-LLaMA3-RM-v0.1, RLHFlow/pair-preference-model-LLaMA3-8B, RLHFlow/ArmoRM-Llama3-8B-v0.1, which are SOTA open-source RMs so far (2024 May).

Step 3.1 Data Generation

We have prepared some prompt sets on huggingface.

UltraFeedback RLHFlow/ultrafeedback_iter1, RLHFlow/ultrafeedback_iter2, RLHFlow/ultrafeedback_iter3
RLHFlow/iterative-prompt-v1-iter1-20K, RLHFlow/iterative-prompt-v1-iter2-20K, RLHFlow/iterative-prompt-v1-iter3-20K ...

To accelerate data generation, we use the VLLM. We prepare two ways of using VLLM to inference for a more robust implementation, where you can try them out and choose the one that fits with your environment best. We use LLaMA3-8B as an example.

You may create a test_gen.sh file, and copy the following contents into the file and run ``bash test_gen.sh''.

# First approach: initialize 4 VLLM processes and split the prompt set to the 4 agents
# The generated samples will be stored at output_dir + local_index + ".jsonl

my_world_size=8 # how many gpu you use
infer_model=RLHFlow/LLaMA3-SFT
prompt_dir=RLHFlow/test_generation_2k
mkdir data
output_dir=./data/gen_data

conda activate vllm
CUDA_VISIBLE_DEVICES=0 python ./generation/get_hf2.py --model_name_or_path ${infer_model} --dataset_name_or_path ${prompt_dir} --output_dir ${output_dir} --K 4 --temperature 1.0 --local_index 0 --my_world_size ${my_world_size}  &
CUDA_VISIBLE_DEVICES=1 python ./generation/get_hf2.py --model_name_or_path ${infer_model} --dataset_name_or_path ${prompt_dir} --output_dir ${output_dir} --K 4 --temperature 1.0 --local_index 1 --my_world_size ${my_world_size}  &
CUDA_VISIBLE_DEVICES=2 python ./generation/get_hf2.py --model_name_or_path ${infer_model} --dataset_name_or_path ${prompt_dir} --output_dir ${output_dir} --K 4 --temperature 1.0 --local_index 2 --my_world_size ${my_world_size}  &
CUDA_VISIBLE_DEVICES=3 python ./generation/get_hf2.py --model_name_or_path ${infer_model} --dataset_name_or_path ${prompt_dir} --output_dir ${output_dir} --K 4 --temperature 1.0 --local_index 3 --my_world_size ${my_world_size}  &
CUDA_VISIBLE_DEVICES=4 python ./generation/get_hf2.py --model_name_or_path ${infer_model} --dataset_name_or_path ${prompt_dir} --output_dir ${output_dir} --K 4 --temperature 1.0 --local_index 4 --my_world_size ${my_world_size}  &
CUDA_VISIBLE_DEVICES=5 python ./generation/get_hf2.py --model_name_or_path ${infer_model} --dataset_name_or_path ${prompt_dir} --output_dir ${output_dir} --K 4 --temperature 1.0 --local_index 5 --my_world_size ${my_world_size}  &
CUDA_VISIBLE_DEVICES=6 python ./generation/get_hf2.py --model_name_or_path ${infer_model} --dataset_name_or_path ${prompt_dir} --output_dir ${output_dir} --K 4 --temperature 1.0 --local_index 6 --my_world_size ${my_world_size}  &
CUDA_VISIBLE_DEVICES=7 python ./generation/get_hf2.py --model_name_or_path ${infer_model} --dataset_name_or_path ${prompt_dir} --output_dir ${output_dir} --K 4 --temperature 1.0 --local_index 7 --my_world_size ${my_world_size}  &

# then, we merge the 8 datasets into one dataset.
wait
python ./generation/merge_data.py --base_path ${output_dir} --output_dir ./data/gen_data.json --num_datasets ${my_world_size}

We can also use API server to generate new responses.

mkdir data
conda activate vllm

# register the api server
bash ./generation/register_server.sh RLHFlow/LLaMA3-SFT

# start to generate
python ./generation/gen_hf.py --ports 8000 8001 8002 8003 8004 8005 8006 8007 --tokenizer RLHFlow/LLaMA3-SFT --dataset_name_or_path RLHFlow/test_generation_2k --output_dir ./data/gen_data.jsonl --K 4 --temperature 1.0

Step 3.2 Data Annotation

Then, we call the reward/preference model trained in step 2 to rank the generated responses.

accelerate launch ./annotate_data/get_rewards.py --dataset_name_or_path ./data/gen_data.jsonl --output_dir ./data/data_with_rewards.jsonl --K 4

If you encounter error ``TypeError: Got unsupported ScalarType BFloat16'', considering adjusting your transformer version.

Step 3.3 Training

conda activate rlhflow
accelerate launch --config_file ./configs/zero2.yaml dpo_iteration/run_dpo.py ./config/training.yaml

If you encounter ``RuntimeError: CUDA error: invalid device ordinal, CUDA kernel errors might be asynchronously reported at some other API call'', you need to adjust num_of_process in the config file according to your GPUs.

Putting Everything Together

We put everything together so that the iterative training can run automatically. Note that we set sleep 1m to wait for registering the API for inference. You may need to adjust this parameter according to your environment.

bash run_loop.sh

Acknowledgement

The authors would like to thank the great open-source communities, including the Huggingface TRL team, the Huggingface H4 team, the Allen Institute AI RewardBench team, the Meta LLaMA team, and Axolotl team for sharing the models, codes, and training sets.

Citation

If you find the content of this repo useful, please consider cite it as follows:

@misc{dong2024rlhf,
      title={RLHF Workflow: From Reward Modeling to Online RLHF}, 
      author={Hanze Dong and Wei Xiong and Bo Pang and Haoxiang Wang and Han Zhao and Yingbo Zhou and Nan Jiang and Doyen Sahoo and Caiming Xiong and Tong Zhang},
      year={2024},
      eprint={2405.07863},
      archivePrefix={arXiv},
      primaryClass={cs.LG}
}
@inproceedings{xiong2023iterative,
  title={Iterative preference learning from human feedback: Bridging theory and practice for RLHF under KL-constraint},
  author={Xiong, Wei and Dong, Hanze and Ye, Chenlu and Wang, Ziqi and Zhong, Han and Ji, Heng and Jiang, Nan and Zhang, Tong},
  booktitle={ICLR 2024 Workshop on Mathematical and Empirical Understanding of Foundation Models}
}

For Tasks:

Click tags to check more tools for each tasks

train language models fine-tune models generate responses annotate data iterative training

For Jobs:

machine learning engineer data scientist research scientist ai engineer nlp researcher

Alternative AI tools for Online-RLHF

Similar Open Source Tools

Online-RLHF

github

: 386

llm-applications

A comprehensive guide to building Retrieval Augmented Generation (RAG)-based LLM applications for production. This guide covers developing a RAG-based LLM application from scratch, scaling the major components, evaluating different configurations, implementing LLM hybrid routing, serving the application in a highly scalable and available manner, and sharing the impacts LLM applications have had on products.

github

: 1.5k

neural

Neural is a Vim and Neovim plugin that integrates various machine learning tools to assist users in writing code, generating text, and explaining code or paragraphs. It supports multiple machine learning models, focuses on privacy, and is compatible with Vim 8.0+ and Neovim 0.8+. Users can easily configure Neural to interact with third-party machine learning tools, such as OpenAI, to enhance code generation and completion. The plugin also provides commands like `:NeuralExplain` to explain code or text and `:NeuralStop` to stop Neural from working. Neural is maintained by the Dense Analysis team and comes with a disclaimer about sending input data to third-party servers for machine learning queries.

github

: 417

depthai

This repository contains a demo application for DepthAI, a tool that can load different networks, create pipelines, record video, and more. It provides documentation for installation and usage, including running programs through Docker. Users can explore DepthAI features via command line arguments or a clickable QT interface. Supported models include various AI models for tasks like face detection, human pose estimation, and object detection. The tool collects anonymous usage statistics by default, which can be disabled. Users can report issues to the development team for support and troubleshooting.

github

: 927

linkedin-api

The Linkedin API for Python allows users to programmatically search profiles, send messages, and find jobs using a regular Linkedin user account. It does not require 'official' API access, just a valid Linkedin account. However, it is important to note that this library is not officially supported by LinkedIn and using it may violate LinkedIn's Terms of Service. Users can authenticate using any Linkedin account credentials and access features like getting profiles, profile contact info, and connections. The library also provides commercial alternatives for extracting data, scraping public profiles, and accessing a full LinkedIn API. It is not endorsed or supported by LinkedIn and is intended for educational purposes and personal use only.

github

: 1.8k

LafTools

LafTools is a privacy-first, self-hosted, fully open source toolbox designed for programmers. It offers a wide range of tools, including code generation, translation, encryption, compression, data analysis, and more. LafTools is highly integrated with a productive UI and supports full GPT-alike functionality. It is available as Docker images and portable edition, with desktop edition support planned for the future.

github

: 309

mint-bench

MINT benchmark aims to evaluate LLMs' ability to solve tasks with multi-turn interactions by (1) using tools and (2) leveraging natural language feedback.

github

: 84

air-light

Air-light is a minimalist WordPress starter theme designed to be an ultra minimal starting point for a WordPress project. It is built to be very straightforward, backwards compatible, front-end developer friendly and modular by its structure. Air-light is free of weird "app-like" folder structures or odd syntaxes that nobody else uses. It loves WordPress as it was and as it is.

github

: 945

gpt-researcher

GPT Researcher is an autonomous agent designed for comprehensive online research on a variety of tasks. It can produce detailed, factual, and unbiased research reports with customization options. The tool addresses issues of speed, determinism, and reliability by leveraging parallelized agent work. The main idea involves running 'planner' and 'execution' agents to generate research questions, seek related information, and create research reports. GPT Researcher optimizes costs and completes tasks in around 3 minutes. Features include generating long research reports, aggregating web sources, an easy-to-use web interface, scraping web sources, and exporting reports to various formats.

github

: 14.5k

contoso-chat

Contoso Chat is a Python sample demonstrating how to build, evaluate, and deploy a retail copilot application with Azure AI Studio using Promptflow with Prompty assets. The sample implements a Retrieval Augmented Generation approach to answer customer queries based on the company's product catalog and customer purchase history. It utilizes Azure AI Search, Azure Cosmos DB, Azure OpenAI, text-embeddings-ada-002, and GPT models for vectorizing user queries, AI-assisted evaluation, and generating chat responses. By exploring this sample, users can learn to build a retail copilot application, define prompts using Prompty, design, run & evaluate a copilot using Promptflow, provision and deploy the solution to Azure using the Azure Developer CLI, and understand Responsible AI practices for evaluation and content safety.

github

: 431

autoarena

AutoArena is a tool designed to create leaderboards ranking Language Model outputs against one another using automated judge evaluation. It allows users to rank outputs from different LLMs, RAG setups, and prompts to find the best configuration of their system. Users can perform automated head-to-head evaluation using judges from various platforms like OpenAI, Anthropic, and Cohere. Additionally, users can define and run custom judges, connect to internal services, or implement bespoke logic. AutoArena enables users to run the application locally, providing full control over their environment and data.

github

: 65

NekoImageGallery

NekoImageGallery is an online AI image search engine that utilizes the Clip model and Qdrant vector database. It supports keyword search and similar image search. The tool generates 768-dimensional vectors for each image using the Clip model, supports OCR text search using PaddleOCR, and efficiently searches vectors using the Qdrant vector database. Users can deploy the tool locally or via Docker, with options for metadata storage using Qdrant database or local file storage. The tool provides API documentation through FastAPI's built-in Swagger UI and can be used for tasks like image search, text extraction, and vector search.

github

: 58

gemini-api-quickstart

This repository contains a simple Python Flask App utilizing the Google AI Gemini API to explore multi-modal capabilities. It provides a basic UI and Flask backend for easy integration and testing. The app allows users to interact with the AI model through chat messages, making it a great starting point for developers interested in AI-powered applications.

github

: 128

IOPaint

IOPaint is a free and open-source inpainting & outpainting tool powered by SOTA AI model. It supports various AI models to perform erase, inpainting, or outpainting tasks. Users can remove unwanted objects, defects, watermarks, or people from images using erase models. Additionally, diffusion models can replace objects or perform outpainting. The tool also offers plugins for interactive object segmentation, background removal, anime segmentation, super resolution, face restoration, and file management. IOPaint provides a web UI for easy access to the latest AI models and supports batch processing of images through the command line. Developers can contribute to the project by installing front-end dependencies, setting up the backend, and starting the development environment for both front-end and back-end components.

github

: 18.7k

agents

Polymarket Agents is a developer framework and set of utilities for building AI agents to trade autonomously on Polymarket. It integrates with Polymarket API, provides AI agent utilities for prediction markets, supports local and remote RAG, sources data from various services, and offers comprehensive LLM tools for prompt engineering. The architecture features modular components like APIs and scripts for managing local environments, server set-up, and CLI for end-user commands.

github

: 60

TaskWeaver

TaskWeaver is a code-first agent framework designed for planning and executing data analytics tasks. It interprets user requests through code snippets, coordinates various plugins to execute tasks in a stateful manner, and preserves both chat history and code execution history. It supports rich data structures, customized algorithms, domain-specific knowledge incorporation, stateful execution, code verification, easy debugging, security considerations, and easy extension. TaskWeaver is easy to use with CLI and WebUI support, and it can be integrated as a library. It offers detailed documentation, demo examples, and citation guidelines.

github

: 5.3k

For similar tasks

argilla

Argilla is a collaboration platform for AI engineers and domain experts that require high-quality outputs, full data ownership, and overall efficiency. It helps users improve AI output quality through data quality, take control of their data and models, and improve efficiency by quickly iterating on the right data and models. Argilla is an open-source community-driven project that provides tools for achieving and maintaining high-quality data standards, with a focus on NLP and LLMs. It is used by AI teams from companies like the Red Cross, Loris.ai, and Prolific to improve the quality and efficiency of AI projects.

github

: 3.9k

Online-RLHF

github

: 386

OlympicArena

OlympicArena is a comprehensive benchmark designed to evaluate advanced AI capabilities across various disciplines. It aims to push AI towards superintelligence by tackling complex challenges in science and beyond. The repository provides detailed data for different disciplines, allows users to run inference and evaluation locally, and offers a submission platform for testing models on the test set. Additionally, it includes an annotation interface and encourages users to cite their paper if they find the code or dataset helpful.

github

: 74

reductstore

ReductStore is a high-performance time series database designed for storing and managing large amounts of unstructured blob data. It offers features such as real-time querying, batching data, and HTTP(S) API for edge computing, computer vision, and IoT applications. The database ensures data integrity, implements retention policies, and provides efficient data access, making it a cost-effective solution for applications requiring unstructured data storage and access at specific time intervals.

github

: 170

mindsdb

MindsDB is a platform for customizing AI from enterprise data. You can create, serve, and fine-tune models in real-time from your database, vector store, and application data. MindsDB "enhances" SQL syntax with AI capabilities to make it accessible for developers worldwide. With MindsDB’s nearly 200 integrations, any developer can create AI customized for their purpose, faster and more securely. Their AI systems will constantly improve themselves — using companies’ own data, in real-time.

github

: 26.6k

training-operator

Kubeflow Training Operator is a Kubernetes-native project for fine-tuning and scalable distributed training of machine learning (ML) models created with various ML frameworks such as PyTorch, Tensorflow, XGBoost, MPI, Paddle and others. Training Operator allows you to use Kubernetes workloads to effectively train your large models via Kubernetes Custom Resources APIs or using Training Operator Python SDK. > Note: Before v1.2 release, Kubeflow Training Operator only supports TFJob on Kubernetes. * For a complete reference of the custom resource definitions, please refer to the API Definition. * TensorFlow API Definition * PyTorch API Definition * Apache MXNet API Definition * XGBoost API Definition * MPI API Definition * PaddlePaddle API Definition * For details of all-in-one operator design, please refer to the All-in-one Kubeflow Training Operator * For details on its observability, please refer to the monitoring design doc.

github

: 1.6k

helix

HelixML is a private GenAI platform that allows users to deploy the best of open AI in their own data center or VPC while retaining complete data security and control. It includes support for fine-tuning models with drag-and-drop functionality. HelixML brings the best of open source AI to businesses in an ergonomic and scalable way, optimizing the tradeoff between GPU memory and latency.

github

: 331

nntrainer

NNtrainer is a software framework for training neural network models on devices with limited resources. It enables on-device fine-tuning of neural networks using user data for personalization. NNtrainer supports various machine learning algorithms and provides examples for tasks such as few-shot learning, ResNet, VGG, and product rating. It is optimized for embedded devices and utilizes CBLAS and CUBLAS for accelerated calculations. NNtrainer is open source and released under the Apache License version 2.0.

github

: 135

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.

github

: 7.1k

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.

github

: 390

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.

github

: 159

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.

github

: 572

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.

github

: 27.7k

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

github

: 897

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.

github

: 951

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.

github

: 2.0k