Quick Links: Documentation | FAQ | Hugging Face | Calendar | Slack | YouTube | X
InstructLab is an approachable open source AI community project. Our community's mission is to enable anyone to shape the future of generative AI via the collaborative improvement of open source-licensed Granite large language models (LLMs) using InstructLab's fine-tuning technology.
Have you ever asked an LLM a question about a subject you know well, and it gave an incorrect answer? “If only I could submit a patch to fix it,” you may have thought. With InstructLab, you can!
InstructLab allows anyone to improve an existing LLM by fine-tuning it with additional data sources. This allows LLMs to continuously gain new knowledge, supplementing gaps in their initial training, even about current events that happened since their pre-training phase.
Subject matter experts from any domain—for example, archaeology, astronomy, and human anatomy—can use InstructLab to teach the Granite LLM family new information.
Project users can experiment with making updates to a quantized version of the Granite model and then rebuild that version locally, checking to see if the update improved the quality of the model’s responses. InstructLab’ s tools allow contributors to both see preview of a newly tuned model, and also check to ensure that a submission to improve the community model is properly composed.
You can learn more about InstructLab and how it works in the project documentation.
Check out InstructLab on Hugging Face.
- Install InstructLab locally
- Contribute to the InstructLab community model
- Contribute to the CLI project code
- Connect with the community to share feedback and ideas, ask questions, and troubleshoot issues.
InstructLab uses Large-Scale Alignment for ChatBots [1] (LAB), a new alignment tuning method for LLMs that leverages synthetic data. To learn more about InstructLab’s origins, visit the About Taxonomy page.
[1] Shivchander Sudalairaj*, Abhishek Bhandwaldar*, Aldo Pareja*, Kai Xu, David D. Cox, Akash Srivastava*. "LAB: Large-Scale Alignment for ChatBots", arXiv preprint arXiv: 2403.01081, 2024. (* denotes equal contributions)
