laion-ai/open-assistant

Open-Assistant was a research project by LAION-AI that aimed to build an open-source chat assistant similar to ChatGPT, one that anyone could run, study, or improve. The project is now completed and no longer actively developed, but its final dataset (oasst2) is publicly available on HuggingFace. The problem it addressed was that capable conversational AI was locked inside proprietary systems, out of reach for researchers and developers who wanted to study or extend it. The project worked in three stages inspired by the InstructGPT research paper. First, the community crowdsourced a large set of human-written instruction and response pairs, essentially, people submitting good examples of what a helpful AI should say. Second, those examples were used to train a reward model that could judge whether a given AI response was good or bad. Third, that reward model was used to fine-tune a language model through reinforcement learning, teaching it to give responses that humans rate highly. Contributors helped by chatting with the AI and giving thumbs-up or thumbs-down ratings to its answers. You would reference this project if you were a researcher wanting to understand how instruction-tuning and human feedback training work in practice, or if you wanted to use the oasst2 dataset to train your own conversational model. The project's architecture used a Python backend, a Next.js web frontend for the data collection and chat interface, and PostgreSQL for storage. Everything was packaged with Docker so contributors could run the full stack locally. The primary language is Python, with Next.js handling the web layer.