// the find
LAION-AI/Open-Assistant
OpenAssistant is a chat-based assistant that understands tasks, can interact with third-party systems, and retrieve information dynamically to do so.
Open-Assistant was LAION's attempt to build an open-source ChatGPT alternative using the InstructGPT RLHF pipeline — crowdsourced data collection, reward model training, and fine-tuning. The project is officially finished and archived; its main deliverable is the oasst2 dataset on HuggingFace, not a running system you can deploy.
The oasst2 dataset is the real artifact here and it's genuinely useful — multilingual, human-annotated conversation trees with ranking data, which is hard to find at this scale. The backend data pipeline is well-structured: Alembic migrations, clear separation between task/message/user models, and a thoughtful tree-based conversation schema that captures branching dialogue properly. The crowdsourcing infrastructure for collecting ranked completions was non-trivial and the code reflects real engineering effort on rate limiting, troll detection, and leaderboard mechanics.
The project is dead — last push August 2024, officially declared complete in October 2023, and the inference system was never production-ready anyway. The monorepo is enormous and disorganized: 30+ dataset scripts in `data/datasets/` that are one-off Jupyter notebooks with no shared tooling. If you wanted to actually run this locally for inference, the docs punt you to a separate `inference/` folder and warn you it's not meant for local use. The RLHF training code isn't even in this repo — it lived in separate projects, so what you're cloning is mostly the data collection UI and backend, not a working model pipeline.