git2llm cleans Git history for model fine-tuning

git2llm is a command-line tool and Python library that turns GitHub repositories into training data for large language models. A developer’s commit history can show how they write, think through, and fix code, but raw commits are messy. They often include merge commits, bot-made changelogs, throwaway work, unclear messages, auto-generated files, and repeated code changes across branches.

git2llm collects commits, pull requests, and issues from public or private repositories in parallel. It then runs a four-step cleanup process that removes merge commits and bot noise, filters work-in-progress and auto-generated content, and removes near-duplicate changes. The near-duplicate check uses MinHash LSH, which finds very similar items instead of only exact copies.

The output can be saved in Alpaca or ShareGPT format and used with tools such as Unsloth, LLaMA-Factory, or other supervised fine-tuning pipelines. In one set of personal repositories, the cleanup removed 78% of the raw commits.