| { | |
| "dataset": { | |
| "name": "FSD50K", | |
| "source_part": "part_1", | |
| "total_samples_evaluated": 8194, | |
| "classes_detected": 199 | |
| }, | |
| "model": { | |
| "checkpoint": "best_fsd50k_primitive.pt", | |
| "architecture": "PrimitiveExtractor (Conv1d + LayerNorm + MultiheadAttention)", | |
| "pretrained_base": "Clotho" | |
| }, | |
| "metrics": { | |
| "mAP_macro_percent": 61.25, | |
| "micro_f1_score_percent": 56.63, | |
| "macro_f1_score_percent": 43.46, | |
| "top_1_hit_rate_percent": 80.84 | |
| }, | |
| "analysis": { | |
| "mAP_note": "Higher than validation-only mAP (52.72%) because it includes training data and rare classes that were missing from the small validation split.", | |
| "macro_f1_note": "Lower than Micro F1, indicating the model still has room to improve on the rarest classes (long-tail distribution).", | |
| "top_1_note": "Excellent reliability. The model's highest confidence guess is a correct label 8 out of 10 times." | |
| }, | |
| "metadata": { | |
| "timestamp": "2026-10-06T09:10:26.818340", | |
| "date": "October 06, 2026", | |
| "repo": "Hitelcy/WindoneRepo2" | |
| } | |
| } |