Spaces:
Running on Zero
Running on Zero
Commit History
2 days ago
-
Fix BRCA training collapse with lab-grid defaults
Training on BRCA with the previous defaults (Adam, lr 1e-3, weight decay 0.01) collapsed: every sample got the same prediction (macro F1 0.133). - Default to the configuration from the Bozdag Lab grid that was stable across 8 seeds on BRCA: Adam, lr 0.01, L2 0.01, batch 32, projection 64, tanh/relu/tanh activations, 200 epochs (macro F1 0.749 +/- 0.018). - Make the three activation functions and SGD momentum (default 0.03) configurable in Hyperparameters. model.py and the artifact format are unchanged; artifacts already store their activations. - Make the automatic GenePT choice the lab-recommended single-cell file. - Default to a 75/25 stratified train/validation split, with helper text and a clear error when a class is too small to stratify. - Fix the Results page training-history plot, which was never populated. - Add stable element ids for e2e tests and a layout tweak for short desktop screens. - Tests: split, leakage, class weights, history plot, activations, moment…9e61493
-
Fix mobile dashboard layout
9a10d98
-
Redesign workbench dashboard shell (#1) Running
- Redesign workbench dashboard shell (72c0d0ba7a7859042da2633a5c048ec50f59cd87) - Fix dataset loading and add intro cards (121eaa8832a7dfe6e03d01b13909bd13d51934fe) Co-authored-by: Faruk Yildirim <F4RUKYLDRM@users.noreply.huggingface.co>2e6a2a4