OpenAssistant has made its models, training data, and code available as part of an effort to build an open-source AI assistant. The release gives researchers access to both the systems and the conversation material used to train them, while the project’s own paper warns that the models can still produce unsafe or biased results.
A dataset built with volunteer contributions
The project began in December, shortly after OpenAI released ChatGPT. Its goal is to create an open-source assistant with similar capabilities. To support that work, more than 13,500 volunteers helped assemble a dataset of assistant-style conversations.
The collection includes 161,443 messages across 66,497 conversation trees, in 35 different languages. It also includes 461,292 quality ratings. Those ratings offer a way to assess the material, while the conversation trees preserve exchanges as connected discussions rather than isolated messages.
OpenAssistant is releasing this dataset as OpenAssistant Conversations, alongside code and models. The project also provides a web interface where people can try the models, evaluate conversations, and contribute feedback that can help improve them.
Several models, with different licensing
The team used the collected instructional data to refine language models, including versions based on Meta's LLaMA model and EleutherAI's Pythia model. The largest announced version is based on LLaMA and has 30 billion parameters.
These are instruction-tuned models: they have been adapted to follow instructions. The source says they have not been further improved through reinforcement learning with human feedback, or RLHF. A comparative study with volunteers suggested that chatbot results could approach those of ChatGPT's gpt-3.5-turbo model, though that comparison does not remove the limitations the project describes.
The Pythia models are already available, while the LLaMA models were expected to be released soon. Their licensing differs: Pythia models are licensed for commercial use, while Meta's licensing means the LLaMA models cannot be used commercially.
The team says it is experimenting with plugins, including Google search, and plans to train and release a LLaMA-30B model with RLHF in the future. These are ongoing plans, separate from the models and materials already released.
Openness comes with safety limits
The accompanying paper describes familiar large language model problems, including hallucinations. It also points to the makeup of the people who contributed annotations: most were male, and the median age was 26. The authors warn that this profile may shape the dataset’s values, perspectives, and interests, which could in turn affect model behavior.
The team has tried to identify and remove harmful messages from the dataset, but says that process is not infallible. The paper recommends using the models in academic research contexts only and urges researchers to investigate their safety and bias before using them in other tasks. It also cautions that the models may behave unsafely and are likely susceptible to prompt injection attacks.
That caution matters for anyone interpreting the release. Open access makes it possible to inspect and study the models and their data, but it does not establish that their outputs are reliable or safe for every use.
Broadening access to alignment research
OpenAssistant frames the release as a way to widen participation in research on large language models and alignment, the adaptation of models to human values. The team argues that this work has largely been limited to a small number of research labs able to collect data and train large models.
Publishing models and data is intended to let more people study alignment challenges. The project also positions this approach against increasingly opaque development and data sourcing, and alignment research conducted by a small group of selected specialists.
OpenAssistant was founded by Andreas Köpf, Yannic Kilcher, Huu Nguyen, and Christoph Schumann. It includes a team of over 20 developers, data and security experts, and a moderation and documentation team. Supporters include Redmond AI, Hugging Face, Weights & Biases, Stabilty AI, and LAION.