HawkFranklin-Research/pelliscope-siglip2-mil

Version status — historical 20 September 2026 upload

These weights are an earlier production snapshot uploaded on 20 September 2026. They are not the final 27 September 2026 ten-seed (42–51) SigLIP2 MIL ensemble with validation-fitted per-condition Platt calibration used for the revised manuscript's reported results (macro ROC-AUC approximately 0.851; macro average precision approximately 0.304 on the locked 750-case cohort).

Do not use this upload to claim reproduction of those final ensemble results. Final-model checkpoint availability remains pending; the code release and saved final predictions are available in pelliscope-train. No checkpoint weights or historical result files were changed by this notice.

Pelliscope SigLIP-2 Gated-Attention MIL (25-Class Teledermatology)

Historical 20 September 2026 model checkpoint for the 25-condition case-level teledermatology Multiple Instance Learning (MIL) model from HawkFranklin Research.

Architecture

• Backbone: SigLIP-2 SO400M (image-level 1152-d frozen embeddings)
• Aggregation: Case-level Gated Attention Multiple Instance Learning (Ilse et al.)
• Conditions: 25 top dermatological conditions from the SCIN dataset
• Input: Bag of 1 to 3 clinic/self-captured skin images per encounter

Included Files

• model.pt: Trained PyTorch Gated Attention MIL weights
• validation_selected_thresholds.csv: Validation-tuned operating thresholds per condition
• test_operating_points.csv: Out-of-sample test sensitivity, specificity, and F1 per condition
• mil_config.yaml: Architecture hyperparameters and training configuration