AI Lab · by Hana Choi · 2026-10-14

Fine-tuned a small open model on my own commit history — it now writes commit messages in my exact style

Took about 3 years of git log output, cleaned it up into instruction pairs, and did a lightweight LoRA finetune on a 7b model. Now it suggests commit messages that genuinely sound like me, abbreviations and all. Took one afternoon end to end on a single consumer gpu. Happy to answer questions about the pipeline.

0 sparks · 9 comments

Comments

Curious Baker · 0 likes

this is such a fun small project, did you worry at all about leaking anything sensitive from commit messages into the weights?

Quiet Fern · 0 likes

good question, i scrubbed anything with ticket numbers or internal hostnames before training just to be safe

Curious Baker · 0 likes

how many examples did you end up with after cleaning? curious about the minimum viable dataset size here

Quiet Runner · 0 likes

around 4000 pairs, definitely could go smaller and still get the style transfer

Aarav Sharma · 0 likes

love this as a proof of concept, could see teams doing this for PR descriptions too

Isha Menon · 0 likes

what gpu are we talking, i want to replicate this this weekend

Jisoo Park · 0 likes

a 3090 was plenty for a 7b lora, dont need anything crazy for this scale

Lucía García · 0 likes

next step: finetune it on your code review comments and have it roast your own prs

Midnight Atlas · 0 likes

this made me laugh way too hard, please do this