Spaces:
Sleeping
Sleeping
Commit History
29 days ago
-
Remove em dashes from the Space
e11747c
-
Serve the v2 locked-test model and report its AUC
The Space advertised "AUC 0.9983 (Top 1% Kaggle)". The project was never submitted to Kaggle, and 0.9983 is the v1 validation AUC that drove early stopping, not a held-out figure. It also loaded models/best_model_epoch50_auc0.9980.pth — an epoch-50 save from the v1 run — so simply relabelling it 0.9979 would have attached the v2 locked-test metric to v1 weights. Switch to final_model_v2_locked.pth (SHA256 b08447ab760d6e0f…), the model the report evaluates. Identical 198 tensor keys; it is a bare state_dict, already handled by the existing load branch. Description now reports locked-test AUC 0.9979 (95% CI 0.9975-0.9982, n = 33,004) and the measured ~0.8 s per-patch latency on cpu-basic. predict() and the TTA logic are unchanged. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>af989b2
-
Improve Space card: model, performance, usage, RUO
Verified signatureb6d4383
-
Add TTA x5 to predict() (matches evaluate.py, AUC 0.9983)
775962b
-
fix: correct ResidualHead forward pass to match training architecture
The forward() method was different from the training code in model.py. Missing GELU and Dropout after residual connection caused random predictions. Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>6f3c538
-
Fix: replace deprecated allow_flagging with flagging_mode
79457e5 -
Fix ResidualHead to match checkpoint keys (fc_out, shortcut)
72e39ab -
Add Gradio inference app
cff0429 -
initial commit
Verified signature262b50f