commit-msg-llm

AI-powered commit message generator. Fine-tuned Qwen2.5-0.5B, served from a full MLOps stack.

🌍 Live Demo →

🎨 https://commit-msg.durak.dev        Frontend (Next.js)
🔌 https://commit.durak.dev            API (FastAPI)
📚 https://commit.durak.dev/docs       OpenAPI docs
📊 https://commit-grafana.durak.dev    Grafana dashboard
🤗 huggingface.co/biighunter/commit-msg-llm-adapter

---

What it does

Paste a git diff → get a clean, conventional commit message.

Powered by Qwen2.5-0.5B fine-tuned with LoRA on 6,676 real commits from the CommitChronicle dataset.

Stack

Layer	Tech
Training	Qwen2.5-0.5B · LoRA · PEFT · Kaggle GPU
Tracking	MLflow
Registry	HuggingFace Hub · ghcr.io
Serving	FastAPI · Redis cache · PyTorch
Infra	Kubernetes (DigitalOcean) · Helm
Ingress	Nginx · cert-manager · Let's Encrypt
Monitoring	Prometheus · Grafana · custom metrics
CI/CD	GitHub Actions · release-please
Eval	BLEU · ROUGE · GPT-4o-mini LLM-as-judge
Frontend	Next.js 15 · TypeScript · Tailwind · shadcn/ui · Framer Motion

Architecture

                                ┌──────────────────┐
                                │  GitHub Actions  │
                                │   (test→build→   │
                                │   push→deploy)   │
                                └─────────┬────────┘
                                          │
                                          ▼
┌─────────┐      HTTPS       ┌──────────────────────────┐
│ Browser │ ─────────────────│  K8s cluster (DO fra1)   │
└─────────┘                  │                          │
                             │  ┌────────┐  ┌─────────┐ │
                             │  │ web    │──│  api    │ │
                             │  │ Next.js│  │ FastAPI │ │
                             │  └────────┘  └────┬────┘ │
                             │                   │      │
                             │              ┌────▼────┐ │
                             │              │  redis  │ │
                             │              └─────────┘ │
                             │                          │
                             │   Prometheus → Grafana   │
                             └──────────────────────────┘

Run locally

# Backend
make install
make serve  # http://localhost:8000

# Frontend
cd web && npm install && npm run dev  # http://localhost:3000

Training

make data    # download CommitChronicle (~10k samples)
make train   # LoRA fine-tune (GPU recommended)
make eval    # BLEU + ROUGE + LLM judge

Deployment

Push to main → GitHub Actions:

1. Runs tests + lint
2. Builds Docker image
3. Pushes to ghcr.io
4. Rolls out to K8s
5. Smoke-tests health endpoint

Metrics

Custom Prometheus metrics exposed at /metrics:

• generate_total{status} — total generations by status
• generate_latency_seconds — inference latency histogram
• cache_events_total{event} — Redis cache hits/misses

Visualize at commit-grafana.durak.dev.

License

MIT