Fine-tuned 1.5B model on free Kaggle T4 reviews Node.js code without hallucinating bugs
jeeva1398 · reddit · 2026-10-06
A developer fine-tuned Qwen2.5-Coder-1.5B for fully offline Node.js code review, released as an npm package.
- Training: QLoRA (Unsloth, r=16, 2 epochs, responses-only loss), Q4KM at 986MB, done in 9 minutes on a free Kaggle T4.
- Data: 740 self-built examples — 68 crash types with real parsed stack traces, 74 before/after review scenarios (half clean diffs to teach "No issues found"), plus npm audit reports.
- Results: on 9 clean diffs the base model hallucinated bugs in all 9; the fine-tune flagged none. On buggy diffs, real-bug rate went 48% → 100%; CPU inference 2x faster (5.9s vs 13.6s); package-version hallucination 93% → 100%. Weaker at explaining unseen error types (68% vs 75%).
- Deployment: CLI uses Ollama if running, else pinned node-llama-cpp with SHA-256-verified GGUF; runs as a GitHub Action reviewing PRs on plain CPU runners. The author is upfront that static checks do the real bug-finding; the fine-tune's job is to confirm without inventing issues.
More from coding & agent
- Garry Tan ports Doom to Paul Graham's Bel LISP in 20 minutes using Opus 5.5 — garrytan · 2026-10-07
- L0pht Hacker Chris Wysopal on Securing AI-Written Code: 'Make It Secure' Isn't a Prompt — WeldPond · 2026-10-07
- Designing evals for contract review agents: ten lawyers, ten redlines — graceisford · 2026-10-07
- Engineer reviews 20 agent-written PRs a day: 'I didn't sign up to be a full-time proofreader' — Specialist_Agent3599 · 2026-10-07
- Researcher unveils RSI paradigm: generic disposable agents plus an evolving knowledge base — yisongyue · 2026-10-07
- KNOWS Benchmark Jointly Evaluates Agents on Search, Tools and Visual Understanding — anmarasovic · 2026-10-07