Nocturne v1.2 teacher — first in the family to beat v1 (core calibrated macro-F1 0.1511 vs 0.1488, mAP 0.1541 vs 0.1498). Multi-hot targets + val-mAP selection; epoch 18/40.
4ad0db1 verified | { | |
| "best_epoch": 18, | |
| "select_metric": "map", | |
| "val_best": 0.15352858863472402, | |
| "test_trainprotocol_thr0.5": { | |
| "macro_f1": 0.0951052515330389, | |
| "micro_f1": 0.3952497507025655, | |
| "map": 0.15312775013932942 | |
| }, | |
| "note": "train-time protocol (fixed 0.5 threshold, one row per file). Release numbers come from benchmark.py." | |
| } |