update dataset card

README.md

@@ -25,17 +25,22 @@ into place:
 
 ```bash
 python fetch_data.py --group deployed   # ~3.7 GB — light probes + agreement gate
-python fetch_data.py --group all        # ~64.5 GB — everything
+python fetch_data.py --group all        # HF caches (~58 GB) + Zenodo raw audio (~6.3 GB)
 ```
 
-## Contents (download groups)
+## Contents
+
+This HF dataset holds the **derived caches** (~58 GB). The **raw audio is not re-hosted here**
+— it is the official challenge dataset and is downloaded from Zenodo by `fetch_data.py` when
+you request `--group heavy`/`all` (see *Provenance* below).
 
 | group | size | files | what it is |
 |---|---|---|---|
 | `deployed` | 3.7 GB | 78 | Perch / harmonic / bg-whitened embeddings (`data/perch/*.npz`), BirdMAE parquet, split metadata — runs the light probes + agreement gate |
 | `repro` | 9.5 GB | 1,920 | `legacy/outputs` + `legacy/final` (checkpoints, `ensemble.json`) — replays the historical leaderboard via `import_runs.py` / `train_gate.py` |
-| `heavy` | 25.9 GB | 272,136 | official raw audio (`data/raw_audio`) + the log-mel streaming cache (`data/feature`) for the MTRCNN/EfficientAT path |
+| `heavy` | 19.6 GB | 756 | the log-mel streaming cache (`data/feature`) for the MTRCNN/EfficientAT path (raw audio comes from Zenodo) |
 | `extras` | 25.3 GB | 14 | additional foundation-model embeddings (Perch / sl-BEATs parquets, token tensors) used by the exploratory `analysis/` scripts |
+| *raw_audio* | *6.3 GB* | *271,380* | *official `Development_data.zip` from Zenodo (not stored here); auto-fetched + extracted to `data/raw_audio`* |
 
 Embeddings are frozen-encoder outputs (Google **Perch**, **BirdMAE**, sl-BEATs) plus
 hand-designed acoustic features (harmonic-comb, background-whitened spectrum); they carry no