Update README.md

README.md

@@ -1,3 +1,18 @@
----
-license: mit
----
+---
+license: mit
+---
+
+# TaikoChartEstimator
+
+MIL-based Taiko chart difficulty estimator that predicts difficulty class and star rating from note charts. The model uses Transformer instance encoders, multi-branch attention MIL pooling, monotonic calibration, and multi-task losses (classification, censored regression, within-song ranking) with optional curriculum scheduling.
+
+- Multi-instance learning over beat-aligned windows with stochastic top-k masking to avoid attention collapse
+- Multi-task objectives with censored regression for boundary stars and ranking loss for within-song monotonicity
+- Transformer instance encoder, multi-branch or gated attention aggregator, monotonic spline/MLP calibrator
+- TensorBoard logging, curriculum scheduling, and HuggingFace checkpoints
+
+Our goals are simple:
+
+1. **Star-Level Granularity**: Move beyond traditional 1-10 integer star ratings to provide continuous sub-star difficulty scores (e.g., 9.3 vs 9.7), offering a more precise difficulty metric.
+2. **High-Difficulty Separation**: Address "10-star inflation" by accurately tiering top-level charts, distinguishing between entry-level 10-star songs and those that significantly exceed the nominal boundary.
+3. **Sectional Interpretability**: Provide section-by-section difficulty analysis to identify which specific segments contribute most to the overall rating, giving clear insights into the chart's complexity.