judgebench SigLIP brand judge v2 (+violation negatives) Part of judgebench: which image QC judges survive optimization pressure? A brand fidelity judge for rhode (beauty brand). A judge = these SigLIP weights + calibration.json (rhode train split centroid + Platt scaling params) in this repo. Base: google/siglip so400m patch14 384 (full SiglipModel , contrastively fine tuned). Training: identical to v1 (same base, loss, epochs, competitor negatives) plus 810 corruption negatives as a third SupCon class — palette + typography families only, generated from train split bases; composition/styling/mood families held out. Report card (judgebench Phase 1, 2,622 item test set): brand AUC 0.975 (vs v1 0.99), logo masking delta ~0, brand dial Spearman 0.45. Trained violation families exploded (palette det@5%FPR 0.06–0.10 → 0.54–0.86; typography → 0.49–0.60); held out families stayed at noise. The pre registered ablation answer was memorize, not generalize : violation training buys detection of exactly the families seen and transfers no general concept of 'wrongness'. You cannot enumerate your way to a safe judge. Scoring Score = Platt calibrated cosine similarity between the image embedding…
We use cookies for essential functionality and analytics. You can accept or reject analytics cookies.Cookie policy