rahulraj1406/forgesight-all-folds

ForgeSight FULL — Complete 5-Fold Ensemble

The complete ForgeSight model. Every verdict is produced by all 15 trained checkpoints — fine-tuned DINOv2-Base (518px), SigLIP-So400m (224px) and CLIP ViT-L/14 (224px), 5 cross-validation folds each — plus live ELA / FFT / Wavelet forensic heuristics and a DINOv2 gradient-saliency map.

Built for the DCU 2026 Adv. ML Challenge ("Orthogonal Manifold Fusion"). Paper result for this ensemble (+ isotonic calibration + TTA): F1 = 0.9325 · AUC = 0.9783.

• —Weights: private repo rahulraj1406/forgesight-v11-full-ensemble (15 checkpoints, ~15.3 GB).
• —Per-model score = mean of its 5 fold probabilities; fusion DINOv2 0.45 + SigLIP 0.30 + CLIP 0.25, then DL×0.80 + Forensic×0.20, threshold 0.47.
• —Fold weights are swapped into one resident module per architecture (torch.load(mmap=True) + load_state_dict(assign=True)) so free-tier RAM is never exceeded.
• —A verdict takes ~2–4 minutes on free CPU. For fast (~15 s) fold-0 inference, use the companion demo: rahulraj1406/forgesight.

Requires the HF_TOKEN Space secret with read access to the private weights repo.