Open Assistant Project Detailed Introduction
Project Overview
Open Assistant is an open-source chat AI assistant project developed by the LAION-AI organization. The project aims to provide everyone with access to excellent chat-based large language models (LLMs), creating innovation in language technology through open source.
⚠️ Important Notice: The OpenAssistant project is complete and has ended. The final released dataset can be found at OpenAssistant/oasst2 on HuggingFace.
Project Vision
Open Assistant believes that open-source collaboration can create a revolution in the field of language technology, just as Stable Diffusion has helped the world create art and images in new ways. The ultimate goal of the project is not just to replicate ChatGPT, but to build the assistant of the future, capable of:
- Writing emails and cover letters
- Performing meaningful work
- Using APIs
- Dynamically researching information
- Supporting personalization and expansion
Technical Approach
Core Technology Stack
The main goal of the project is to have a chatbot that can answer questions and better follow instructions by adapting large language models (LLMs). To this end, the project used the method proposed in the InstructGPT paper, which is based on Reinforcement Learning from Human Feedback (RLHF).
Three-Step Training Method
The project follows the three-step method outlined in the InstructGPT paper:
Step 1: Data Collection
- Collect high-quality human-generated instruction-completion samples (prompt + response)
- Goal: Over 50,000 samples
- Design a crowdsourcing process to collect and review prompts
- Avoid training flood attacks/toxic/garbage/personal information data
- Incentivize the community through leaderboards, showcasing progress and most active users
Step 2: Ranking Collection
- Sample multiple completions for each collected prompt
- Randomly present the completions of the prompt to users for ranking (from best to worst)
- Handle unreliable or malicious users through crowdsourcing
- Collect votes from multiple independent users to measure overall consistency
- Use the collected ranking data to train a reward model
Step 3: RLHF Training
- Perform Reinforcement Learning from Human Feedback training phase based on prompts and reward models
- The resulting model can be used to continue the sampling step for the next iteration
Project Architecture
Development Environment Setup
The project supports a complete Docker stack deployment, including the website, backend, and related dependent services.
Basic startup command:
docker compose --profile ci up --build --attach-dependencies
MacOS M1 chip users need to use:
DB_PLATFORM=linux/x86_64 docker compose ...
Local access:
Development Container Support
The project provides standardized development environment support:
- Local VSCode devcontainer
- GitHub Codespaces web browser environment
- Configuration files are located in the
.devcontainer folder
Features
Chat Functionality
- The chat frontend is online, and users can log in and start chatting
- Supports liking or disliking feedback on assistant responses
- Real-time interactive experience
Data Collection
- The data collection frontend is online, and users can log in and start performing tasks
- Directly help improve Open Assistant capabilities by submitting, ranking, and labeling model prompts and responses
- Crowdsourced approach to collecting high-quality data
Inference System
The project includes a complete inference system, supporting:
- Local deployment of inference services
- Ability to run on consumer-grade hardware
- Scalable architecture design
Open Source Features
Community Participation
- The project is organized by LAION and individuals around the world who are interested in bringing this technology to everyone
- Developers are welcome to contribute code
- Detailed contribution guidelines are provided
Project Status
Important Reminder: The OpenAssistant project is complete and has ended. Although the project itself has ended:
- The final dataset oasst2 is available on HuggingFace
- The code is still open source and accessible
- The community can continue to develop based on existing work
Technical Requirements
Hardware Requirements
- The project vision is to create a large language model that can run on a single high-end consumer-grade GPU
- Supports consumer-grade hardware deployment
- Optimized inference performance
Deployment Options
- Docker containerized deployment
- Local development environment
- Cloud deployment support
- Independent deployment of inference services
Related Resources