Add small-sample callout and Croissant pointer

README.md

@@ -87,7 +87,21 @@ Per-city full sizes: Austin ~127 GB, NY ~101 GB, Dallas ~156 GB, Fort Worth ~134
 
 ---
 
-## Quick start
+## Sample for reviewers (small subset, ~4.6 GB)
+
+Per NeurIPS Datasets-track guidance, large datasets should provide a smaller sample. The **per-city mini bundle** is exactly that: 2000 training links + 500 validation + 500 test, baked as PyTorch tensors with normalisation statistics — same schema as the full data, ~3,000× smaller per city.
+
+```bash
+# Smallest single download — Austin mini bundle, ~4.6 GB:
+hf download neurips2026citympc/CityMPC \
+  --repo-type dataset \
+  --include "manifests/city_10_austin_3p5_s/{train_2000,val_500,test_500}.* manifests/city_10_austin_3p5_s/norm_stats.json" \
+  --local-dir .
+```
+
+How the sample was created: deterministic seed-42 random subset of links from the filtered training/validation/test splits (size N=2000/500/500), then the same `bake_dataset.py` pipeline used for the full splits stacks normalised tensors into `.pt` files with the identical schema as `train.h5` / `val.h5` / `test.h5`. The same procedure is available for all 5 cities.
+
+## Quick start (full dataset)
 
 ```bash
 pip install huggingface_hub
@@ -96,6 +110,10 @@ hf download neurips2026citympc/CityMPC --repo-type dataset --include "manifests/
 
 For a one-line download + verify experience, use the citympc `scripts/download_hf.py` helper (see the citympc repo).
 
+## Croissant metadata
+
+A valid Croissant 1.1 metadata file is at the dataset repo root: [`croissant.json`](https://huggingface.co/datasets/neurips2026citympc/CityMPC/resolve/main/croissant.json). It contains all FileObject SHA-256 hashes, the RecordSet schema (channel sample fields + norm-stats fields), Responsible AI properties, and provenance. Validated against `mlcroissant` (0 errors, 0 warnings).
+
 ---
 
 ## Data schema