Model
Qwen2.5-Coder-0.5B-Instruct fine-tuned with LoRA on conventional commit examples from CommitBench. Training reached checkpoint-2000, about 71% of the planned run, before Colab GPU quota cut it off.
A local-first wrapper for a fine-tuned commit message model. Paste a
staged diff, generate a conventional commit subject, then use the
same flow from the terminal with commitgen commit.
Preview is stable for the same diff and updates when the diff changes.
Qwen2.5-Coder-0.5B-Instruct fine-tuned with LoRA on conventional commit examples from CommitBench. Training reached checkpoint-2000, about 71% of the planned run, before Colab GPU quota cut it off.
The tuned adapter learns the commit type prefix instead of producing generic chat responses, moving from 0.8% type accuracy on the base model to 58.2%.
The merged model is quantized to GGUF and served by Ollama, then wrapped by a zero-dependency CLI.
The hosted page is the quick demo. The local install runs the real
quantized model through Ollama and adds a commitgen
command that reads staged Git diffs. Source code:
github.com/Nilay-Mehta/commitgen.
commitgen-Q4_K_M.gguf next to deployment/Modelfile.ollama create commitgen -f deployment/Modelfile.pip install -e deployment from the repo root.commitgen --help.git add.commitgen msg to preview.commitgen commit for confirm-edit-commit flow.# Full local setup git clone https://github.com/Nilay-Mehta/commitgen.git cd commitgen # Add commitgen-Q4_K_M.gguf to deployment/ ollama create commitgen -f deployment/Modelfile pip install -e deployment git add . commitgen commit
# Optional local web preview python -m http.server 8000 # Open: http://localhost:8000/web/
# CLI modes commitgen msg commitgen commit commitgen push # The CLI always asks before committing, # and supports edit/retry from the prompt.