.png)
2 hours ago
1
We trained an SLM assistants for assistance with commit messages on TypeScript codebases - Qwen 3 model (0.6B parameters) that you can run locally!
First, install Ollama, following the instructions on their website.
Then set up the virtual environment:
or using uv:
The model is hosted on huggingface:
Finally, download the models from huggingface and build them locally:
The commit bot with diff the git repository provided via --repository option and suggest a commit message. Use the --watch option to re-run the assistant whenever the repository changes.
The tuned models were trained using knowledge distillation, leveraging the teacher model GPT-OSS-120B. The data+config+script used for finetuning can be found in data. We used 20 typescript git diff examples (created using distillabs' vibe tuning) as seed data and supplemented them with 10,000 synthetic examples across various typescript use cases (frontend, backend, react etc.).
We compare the teacher model and the student model on 10 held-out test examples using LLM-as-a-judge evaluation:
| GPT-OSS (thinking) | 120B | 1.00 |
| Qwen3 0.6B (tuned) | 0.6B | 0.90 |
| Qwen3 0.6B (base) | 0.6B | 0.60 |
Evaluation Criteria: LLM-as-a-judge: The training config file and train/test data splits are available under /data.
Q: Why don't we just use Llama3.X yB for this??
We focus on small models (< 8B parameters), and these make errors when used out of the box (see 5.)
Q: I want to train a small language model for my use-case
A: Visit our website and reach out to us, we offer custom solutions.
