Hi HN,
I have an issue with vague commit messages like "fix PR" or "fix issue". To address this, I’ve been developing a locally fine-tuned model called git-commit-message. The reason I keep refining this model is my belief in Vertical AIs—specialized, lightweight LLMs designed for specific tasks, allowing them to run locally in a secure environment.
Currently, I’m testing a beta version of my model with 3B parameters, comparing it against the initial 7B version.
Inspired by DeepSeek’s reasoning-based approach, I applied a similar technique to git commit messages, and the results have been impressive.
Like any LLM, my model isn't immune to hallucinations, and context size remains a challenge. Larger commits won’t perform as well as smaller ones, but best practices dictate committing small, focused chunks of code anyway.