MCITlib mines GitHub for agent skills
MCITlib is an open-source framework and benchmark suite that automates agent skill acquisition by mining procedural knowledge from existing GitHub repositories. By standardizing extracted code into portable SKILL.md definitions and using continual instruction tuning, it enables Large Multimodal Models to learn specialized tasks without the overhead of full retraining.
MCITlib industrializes the conversion of GitHub repositories into agent capabilities, bridging the gap between static code and dynamic model behavior. It uses dense retrieval and semantic analysis to package procedural skills from open-source repositories and implements over eight continual instruction tuning methods to prevent catastrophic forgetting. The framework reports a 40% efficiency gain in knowledge transfer and provides rigorous benchmarks like UCIT and MLLM-DCL to evaluate learning retention in real-world scenarios.
DISCOVERED
26d ago
2026-03-16
PUBLISHED
26d ago
2026-03-16
RELEVANCE
AUTHOR
Discover AI