Fine-tuned 2B open model hits 67% on turbine defect detection, beating GPT-6's under 10%
helloiamleonie · x · 2026-10-30
Liquid AI's Leonie shares a hands-on demo alongside the LFM3 release: fine-tuning the LFM3-VL-2B vision model to spot wind turbine blade defects in drone photos.
- Baseline: both the untuned LFM3-VL-2B and GPT-6 Luna named the defect correctly in under 10% of images
- Training: 3,000 public photos (WTBs2025 dataset), SFT + GRPO, a few hours on a single A100
- Result: 67% accuracy on 1,400 held-out photos, with more precise localization than GPT-6 Luna, which often mislabeled defect types and produced extra boxes
- Playbook: evaluate whether a frontier model suffices; if not, pick a small open model, curate proprietary examples, fine-tune in hours on one GPU, re-run evals, deploy anywhere
Her argument: frontier models are generalists, but domain-specific problems need specialists — validating the path of fine-tuning open models for customers instead of relying on frontier APIs.
More from coding & agent
- Context Engineering with DSPy: a practical guide to programming—not hand-crafting—LLM context — kieranklaassen · 2026-10-30
- Former games founder protypes an elemental fighting game with 'vibe UI' built on Claude and GPT — trq212 · 2026-10-30
- Proof of Agent rolls out 15-minute checkpoints anchored to Solana and Base, capping validation costs — adamamcbride · 2026-10-30
- Dev updates his site's showcase by @-replying a bot — "build in public" redefined — jonathan_wilke · 2026-10-30
- ~10x speedup for leafcutter as maintainers turn to AI to validate unreviewably large PRs — jmschreiber91 · 2026-10-30
- Your wiki is becoming a codebase: agents will write and review company knowledge — handotdev · 2026-10-30