v2.4: README at sample-subdirectory level explaining how the representative sample was created (per NeurIPS E&D Dataset Large URL guideline)

embeddings/matbench_dielectric/README.md

@@ -0,0 +1,76 @@
+# Representative sample for NeurIPS Evaluations & Datasets review
+
+This directory is the representative sample of the MOOD benchmark
+referenced in the OpenReview submission's "Dataset Large URL" field.
+The full release is ~9.2 GB; per the NeurIPS E&D guidelines for
+datasets > 4 GB, a smaller sample is provided so reviewers can
+inspect data quality without downloading the full archive.
+
+## What's in this directory
+
+Penultimate-layer GNN embeddings for the **`matbench_dielectric`** task
+across all 4 GNN architectures and all 14 retained OOD splits. Each
+`<arch>/<split>/fold_0/` subdirectory contains:
+
+| File | Shape | Description |
+|---|---|---|
+| `train.npy`            | `(n_train,    d)` | Training-pool embeddings (raw, not L2-normalized) |
+| `id_test.npy`          | `(n_id_test,  d)` | ID test-pool embeddings |
+| `ood_test.npy`         | `(n_ood_test, d)` | OOD test-pool embeddings |
+| `train_pred.npy`       | `(n_train,)`      | GNN prediction on training pool |
+| `id_test_pred.npy`     | `(n_id_test,)`    | GNN prediction on ID test |
+| `ood_test_pred.npy`    | `(n_ood_test,)`   | GNN prediction on OOD test |
+| `train_targets.npy`    | `(n_train,)`      | True regression target |
+| `id_test_targets.npy`  | `(n_id_test,)`    | True regression target |
+| `ood_test_targets.npy` | `(n_ood_test,)`   | True regression target |
+
+Embedding dimension `d`: 256 (ALIGNN), 128 (CGCNN), 64 (SchNet), 64 (DimeNet++).
+
+## How this sample was created
+
+The `dielectric` task was selected from the 5 retained MatBench tasks
+because:
+
+1. **Smallest by disk** among method-bottlenecked datasets (~132 MB
+   for embeddings, vs. 328 MB / 339 MB / 32 MB / 328 MB for the others).
+2. **Full benchmark structure preserved**: every GNN architecture and
+   every retained OOD split is represented, so the sample exercises the
+   full evaluation matrix for one task — a reviewer can drop in a new
+   post-hoc detector on any (arch, split) cell here and reproduce the
+   layout used for the headline AUROC table.
+3. **Method-bottlenecked regime**: dielectric is one of the two
+   method-bottlenecked datasets in the benchmark (the other is
+   `log_gvrh`), so a reviewer inspecting data quality sees nontrivial
+   ID/OOD signal in the embeddings rather than the chance-level
+   embeddings characteristic of the encoder-bottlenecked regime.
+
+## Layout in the full release
+
+The other 4 tasks (`matbench_log_gvrh`, `matbench_log_kvrh`,
+`matbench_perovskites`, `matbench_phonons`) follow the identical
+directory structure in the full release. Total full-release size for
+embeddings alone is ~1.1 GB; with checkpoints / flow models / probes /
+results / risk-coverage curves added, the full release is ~9.2 GB.
+
+## Quick verification
+
+```python
+import numpy as np
+from huggingface_hub import hf_hub_download
+
+REPO = "mood-anonymous/mood-benchmark"
+
+z = np.load(hf_hub_download(
+    REPO,
+    filename="embeddings/matbench_dielectric/alignn/comp_halide/fold_0/train.npy",
+    repo_type="dataset",
+))
+print(z.shape, z.dtype)  # (3155, 256) float32
+```
+
+## Full dataset README
+
+For the complete benchmark documentation, intended uses, citation,
+license, and quick-start examples covering all artifact types
+(embeddings, checkpoints, flow models, probes, results, risk-coverage),
+see the [main README](../../README.md) at the dataset root.