Upload folder using huggingface_hub

README.md

@@ -80,3 +80,34 @@ Both inputs should be normalized and have the same temporal dimensions.
 ## Training
 
 The model was trained to detect misaligned or poorly-synchronized rhythm game charts by comparing music-chart pairs with various synthetic corruptions (time shifts, random note placement, etc).
+
+## Evaluation Results
+
+Evaluation was performed on 2,204 test samples with various segment durations. The model uses a severity parameter of 0.56.
+
+### Overall Accuracy by Segment Duration
+
+| Duration | Overall | Positive | Shift | Random | Mismatch |
+|----------|---------|----------|-------|--------|----------|
+| 5s | 81.85% | 95.69% | 79.04% | 97.41% | 97.41% |
+| 10s | 83.35% | 96.55% | 80.60% | 97.41% | 100.00% |
+| 20s | 84.66% | 96.55% | 82.06% | 99.14% | 100.00% |
+| 30s | 85.30% | 95.69% | 82.81% | 100.00% | 100.00% |
+| 60s | 85.98% | 95.69% | 83.62% | 100.00% | 100.00% |
+| 120s | **86.25%** | 94.83% | **84.00%** | 100.00% | 100.00% |
+| 180s | 85.57% | 94.83% | 83.19% | 100.00% | 100.00% |
+
+### Shift Detection by Offset (120s segment)
+
+| Offset | Accuracy | Offset | Accuracy |
+|--------|----------|--------|----------|
+| -0.50s | 91.38% | +0.50s | 92.24% |
+| -0.30s | 89.66% | +0.30s | 89.66% |
+| -0.20s | 91.38% | +0.20s | 94.83% |
+| -0.10s | 100.00% | +0.10s | 100.00% |
+| -0.05s | 100.00% | +0.05s | 100.00% |
+| -0.03s | 91.38% | +0.03s | 95.69% |
+| -0.02s | 84.48% | +0.02s | 88.79% |
+| -0.01s | 20.69% | +0.01s | 13.79% |
+
+> **Note**: The ±10ms shifts are at the edge of perceptual timing tolerance and are intentionally difficult to detect.