maint: update readme, add agents.md

AGENTS.md

@@ -0,0 +1,74 @@
+# AGENTS.md
+
+Guidance for coding agents working in this repository.
+
+## What this repo is
+
+This is the **BeyondArena Datasets** Hugging Face dataset repository — a flat bundle of 142 curated
+tabular datasets used by [TabArena](https://tabarena.ai/). It is **not** a Python package or
+application. There is no build, no tests, no CI pipeline to run here.
+
+The repo is consumed by [Data Foundry](https://github.com/TabArena/data-foundry), which resolves
+each `(unique_name, uuid)` pointer to one of the directories below and loads its six artifact files
+into a `CuratedContainer`. Any change to this repo's layout, naming, or file contents is a breaking
+change for Data Foundry — treat it as a data release, not source code.
+
+## Layout
+
+```
+<dataset_name>/<uuid>/...                 # default (132 datasets)
+<dataset_name>/versions/<uuid>/...        # versioned wrapper (10 large non-IID datasets)
+```
+
+Each leaf `<uuid>/` directory contains exactly six files and nothing else:
+
+```
+dataset.parquet                                       # the table
+dtypes.json                                           # column name → pandas dtype
+container_metadata.json                               # uuid + sha256 checksum
+dataset_metadata.dataset-mold-v1.json                 # provenance & curation notes
+task_metadata.predictive-ml-task-mold-v1.json         # target, problem type, metric, split keys
+experiment_metadata.predictive-ml-splits-mold-v1.json # CV fold indices
+```
+
+Top-level files: `README.md` (the Hugging Face dataset card), `LICENSE`, `dataset_references.bib`,
+and this file.
+
+## Editing rules
+
+- **Do not rename, move, or delete dataset directories or any of the six artifact files inside
+  them.** UUIDs are pinned references from Data Foundry's `final_uuid_list.py`; renaming silently
+  breaks downstream consumers.
+- **Do not regenerate `dataset.parquet`, `container_metadata.json`, or any `*_metadata.*.json` by
+  hand.** They are produced by `CuratedContainer.save()` in Data Foundry. If a dataset needs to
+  change, the change happens in a Data Foundry curation notebook and the new artifact is uploaded
+  here — the checksum in `container_metadata.json` must match the parquet.
+- **The README's YAML frontmatter `configs:` block routes Hugging Face's `datasets` loader to each
+  `dataset.parquet`.** When adding a dataset, add a matching `config_name` entry; preserve the
+  `versions/*/` glob for versioned datasets.
+- **Keep the per-dataset table in the README sorted by `N` (rows).** That's the convention; don't
+  re-sort by name or year.
+- **`dataset_references.bib` must contain one entry per unique `academic_reference_bibtex_key`
+  referenced in the per-dataset `dataset_metadata.dataset-mold-v1.json` files.** Some datasets cite
+  multiple keys — list them comma-separated in the README's `BibKey(s)` column.
+
+## What to do for common requests
+
+- **"Add a dataset"** — don't. New datasets are curated in Data Foundry
+  (`datasets/_dev/<topic>/<unique_name>/<unique_name>.ipynb`), reviewed there, and uploaded to this
+  repo by maintainers. Point the user at
+  [CONTRIBUTING_DATASETS.md](https://github.com/TabArena/data-foundry/blob/main/CONTRIBUTING_DATASETS.md).
+- **"Update a dataset"** — same answer: re-curate in Data Foundry, upload the new UUID directory
+  here, update `configs:` in the README and the per-dataset row.
+- **"Update the README"** — fine. Edit `README.md` directly. The dataset card is rendered on
+  Hugging Face, so preserve the YAML frontmatter and keep markdown compatible with HF's renderer
+  (GitHub-flavored, supports `<details>`).
+- **"Fix a checksum / metadata mismatch"** — investigate the source curation notebook in Data
+  Foundry. Do not edit `container_metadata.json` by hand to match a parquet you regenerated; the
+  parquet itself is wrong.
+
+## Git
+
+The git remote is the Hugging Face dataset hub
+(`huggingface.co/datasets/TabArena/BeyondArena`-style URL). Parquet files are LFS-tracked. Commits
+are visible publicly the moment they're pushed — do not commit unreviewed dataset edits.

README.md

@@ -594,17 +594,13 @@ Datasets from BeyondArena, a unified, holistic benchmark for tabular data that s
 across sample size and feature dimensionality scales, with diverse feature types (with text, with high cardinality) from a
 broad range of disciplines.
 
-We introduce BeyondArena and its datasets in: [TODO link to paper]
+We introduce BeyondArena and its datasets in: [PLACEHOLDER]
 
 <details>
 <summary><b>Click for BibTeX!</b></summary>
 
 ```text
-@article{X,
-  title   = {X},
-  author  = {X},
-  year    = {2026}
-}
+PLACEHOLDER
 ```
 </details>
 
@@ -614,52 +610,45 @@ More details:
 
 ## Quickstart
 
-We suggest using the dataset via DataFoundry ([TODO LINK]):
+We recommend using the datasets via [Data Foundry](https://github.com/TabArena/data-foundry), which
+resolves a curated container (table + dtypes + task metadata + outer CV splits) by name and caches it
+locally:
 
-```python
-# TODO: add code example from DataFoundry here
+```bash
+pip install data-foundry
 ```
 
-### Loading a single dataset directly
-
-Each per-dataset config in this card's frontmatter routes only `dataset.parquet`, which is enough to get
-the table but **not** the sibling metadata files (`dtypes.json`, `task_metadata.*`, `experiment_metadata.*`
-with the CV folds, `dataset_metadata.*`, `container_metadata.json`). Because the benchmark protocol depends
-on those files, the recommended path is to download the whole dataset folder with `huggingface_hub`:
-
 ```python
-from huggingface_hub import snapshot_download
-
-local_dir = snapshot_download(
-    repo_id="TabArena/BeyondArena",
-    repo_type="dataset",
-    allow_patterns=["churn/**"],                # one or more <dataset_name>/** globs
-)
-# local_dir/<dataset_name>/<uuid>/ now contains all six files for that dataset.
-```
-
-For the 10 datasets that use the `versions/` wrapper (see [Dataset Structure](#dataset-structure)), the layout 
-is `<dataset_name>/versions/<uuid>/...` — the `<dataset_name>/**` glob already covers both layouts.
+from data_foundry.collections import BEYOND_ARENA
 
-If you only need the table (no folds, no metadata), the `datasets` library shortcut works:
+container = BEYOND_ARENA.get_dataset("airfoil_self_noise")
+print(container.describe())                  # full identity + dtypes + task + splits
+print(container.dataset.shape)               # the actual DataFrame
+print(container.task_metadata.split_regime)  # "iid", "temporal_non_iid", or "grouped_non_iid"
 
-```python
-from datasets import load_dataset
+df = container.dataset
+target = container.task_metadata.target_column_name
 
-ds = load_dataset("<org>/BeyondArena", name="churn")   # any per-dataset config_name
+for repeat_id, folds in container.experiment_metadata.splits.items():
+    for fold_id, (train_idx, test_idx) in folds.items():
+        X_train, y_train = df.iloc[train_idx].drop(columns=target), df.iloc[train_idx][target]
+        X_test,  y_test  = df.iloc[test_idx].drop(columns=target),  df.iloc[test_idx][target]
+        # ... fit, evaluate ...
 ```
 
-### Downloading the full bundle
+To pre-download the entire collection in a single network round-trip:
 
 ```python
-from huggingface_hub import snapshot_download
+from data_foundry.collections import BEYOND_ARENA
 
-local_dir = snapshot_download(
-    repo_id="<org>/BeyondArena",
-    repo_type="dataset",
-)
+BEYOND_ARENA.prefetch()                          # warms the cache once
+for container in BEYOND_ARENA.iter_containers(): # now hits disk only
+    print(container.dataset_metadata.unique_name, container.dataset.shape)
 ```
 
+See [Data Foundry's examples](https://github.com/TabArena/data-foundry/tree/main/examples) for a full
+benchmarking walkthrough, the three split regimes (IID / temporal / grouped), and the curation flow.
+
 ## Datasets
 
 BeyondArena comes with 142 datasets. BeyondArena covers tabular classification and regression tasks.
@@ -871,6 +860,46 @@ Each directory contains exactly six files:
 
 For details on files and the metadata structure, checkout [DataFoundry](https://github.com/TabArena/data-foundry)!
 
+### Loading a single dataset directly
+
+Each per-dataset config in this card's frontmatter routes only `dataset.parquet`, which is enough to get
+the table but **not** the sibling metadata files (`dtypes.json`, `task_metadata.*`, `experiment_metadata.*`
+with the CV folds, `dataset_metadata.*`, `container_metadata.json`). Because the benchmark protocol depends
+on those files, the recommended path is to download the whole dataset folder with `huggingface_hub`:
+
+```python
+from huggingface_hub import snapshot_download
+
+local_dir = snapshot_download(
+    repo_id="TabArena/BeyondArena",
+    repo_type="dataset",
+    allow_patterns=["churn/**"],                # one or more <dataset_name>/** globs
+)
+# local_dir/<dataset_name>/<uuid>/ now contains all six files for that dataset.
+```
+
+For the 10 datasets that use the `versions/` wrapper (see [Dataset Structure](#dataset-structure)), the layout 
+is `<dataset_name>/versions/<uuid>/...` — the `<dataset_name>/**` glob already covers both layouts.
+
+If you only need the table (no folds, no metadata), the `datasets` library shortcut works:
+
+```python
+from datasets import load_dataset
+
+ds = load_dataset("<org>/BeyondArena", name="churn")   # any per-dataset config_name
+```
+
+### Downloading the full bundle
+
+```python
+from huggingface_hub import snapshot_download
+
+local_dir = snapshot_download(
+    repo_id="<org>/BeyondArena",
+    repo_type="dataset",
+)
+```
+
 ## Licensing
 
 This collection is released under the terms in [LICENSE](LICENSE) (`copyright-at-original-authors`).
@@ -880,10 +909,12 @@ Individual datasets retain their original licenses; see each dataset metadata fo
 
 If you use BeyondArena, please cite:
 
+**PLACEHOLDER**
+
 **BibTeX:**
 
 ```bibtex
-TODO
+PLACEHOLDER
 ```
 
 
@@ -893,4 +924,4 @@ If you use individual datasets, please also cite their original authors. BibTeX
 
 ## Changelog
 
-- **[XXX]** — Initial release: 142 curated IID and non-IID tasks.
+- **[27th May 2026]** — Initial release: 142 curated IID and non-IID tasks.