InstructLab

An new community-based approach to build truly open-source LLMs

A cute dog

Weekly release cycle for models and data

Each week, the InstructLab community model will be updated with the latest contributions and shared on Hugging Face.

Synthetic data generated by the project will also be made available.

InstructLab
Open source community experiments with skills
Contributors open pull requests for InstructLab
New skills merged into community LLM
New skills recipes
Approved pull requests
Updated community model

How it works

detect-gap doesn't exist
1
InstructLab can augment models through skills recipes used to generate synthetic data for tuning. Experiments can be run locally on quantized version of these models.
taxonomy doesn't exist
2
Skills recipes take the form of example inputs / outputs for a desired skill. These skills are organized in a structured taxonomy and anyone can contribute to it.
synthetic-data doesn't exist
3
InstructLab uses the skills recipes to systematically generate new synthetic data, which is also made available to the community.
fine-tune doesn't exist
4
The InstructLab base model is re-tuned using all synthetic data generated to date. This includes any new contributions, which introduce new skills.

Taxonomy

A taxonomy of skills and knowledge helps to identify gaps in desired capabilities, and then generate enough diversity in synthetic data to tune base models efficiently.

$

Join the community

To get started, download the ilab command line interface and a quantized version of the latest model from the community GitHub.

You can experiment locally until your model is producing the desired outputs, and then contribute your skills recipes and knowledge sources back to the community.