Upload folder using huggingface_hub

README.md

@@ -16,7 +16,7 @@ A neural network model for evaluating the quality of rhythm game charts relative
 The model uses an early fusion approach with dilated convolutions for temporal analysis:
 
 1. **Early Fusion**: Concatenates music and chart mel spectrograms along the channel dimension (80 + 80 = 160 channels)
-2. **Dilated Residual Encoder**: 4 residual blocks with increasing dilation rates (1, 2, 4, 8) to capture multi-scale temporal context while preserving 11ms frame resolution
+2. **Dilated Residual Encoder**: 4 residual blocks with increasing dilation rates (1, 2, 4, 8) to capture multi-scale temporal context while preserving 11ms frame resolution. This gives the model a **receptive field of ~0.73s** (63 frames), meaning each time-step's score depends on the local ~0.36s context before and after.
 3. **Error-Sensitive Scoring Head**: Combines average local scores with the worst 10% of scores using a learnable mixing parameter
 
 ```