LoRA fine-tuning fixes 468 misclassified tumour patches — but introduces 217 new errors
burny_tech · x · 2026-11-02
While improving cancer classification models on histopathology patches, the author shared a validation comparison showing LoRA fine-tuning fixed 468 previously misclassified patches, but introduced errors on 217 previously correct ones (1,605 wrong both ways, 30,478 right both ways), benchmarked against the original model, DINOv2-L and a colour baseline.
He also worries that biology-focused guardrails may make Claude unusable for legitimate cancer-classification research — a critique of overly strict safety gating blocking medical AI work.
More from Models
- Developers mourn ChatGPT's unmetered era ending: 'The Meter' kills exploratory flow — doodlestein · 2026-11-02
- Anthropic eats a ~$2B ban loss to protect distillation-driven model moat — oran_ge · 2026-11-02
- Even heavy users can't tell which models power the free tiers of ChatGPT and Claude — iScienceLuvr · 2026-11-02
- Grok to embed invisible watermarks in AI-generated text from Nov 7 to comply with EU AI Act — Polymarket · 2026-11-02
- GPT-6.1 medium in fast mode tested on chess puzzles — imjustnewatai · 2026-11-02
- Gemini 4 sweeps evals: new SOTA on WANDR and Riemann-Bench, 2-3x cheaper Argon — gabriberton · 2026-11-02