Card: list the training data (datasets:, base_model:) and add a Training data section

Follow-up to braindecode PR #1493: `from_pretrained` reads `datasets:` into `model.card_data`. Ids: MOABB names, NEMAR nm/on ids, OpenNeuro ds ids; evidence per checkpoint.

README.md

@@ -1,3 +1,8 @@
+---
+datasets:
+- nm000185
+---
+
 # plot_sleep_staging_eldele2021
 
 Pretrained artifacts for the Braindecode tutorial `examples/applied_examples/plot_sleep_staging_eldele2021.py`.
@@ -9,3 +14,21 @@ These files are meant to be loaded by the tutorial so the docs can show stable p
 - `params.safetensors`: classifier parameters
 - `history.json`: Skorch training history used by the tutorial plots
 - `metadata.json`: summary metrics for the stored checkpoint
+
+<!-- braindecode:training-data -->
+## Training data
+
+The data that shaped these weights. The ids are the `datasets:` field of this card,
+which braindecode reads into `model.card_data.datasets` when you call
+`from_pretrained` (MOABB class names for MOABB datasets, OpenNeuro / NEMAR
+accessions, Hub ids otherwise). `nm`/`on` ids are datasets on [NEMAR](https://nemar.org)
+(`on` = NEMAR's copy of the OpenNeuro dataset with the same number). Rows marked
+`private:` are data the authors did not release; they are not in `datasets:`.
+
+| `datasets:` id | Dataset | Stage | Modality |
+|---|---|---|---|
+| [`nm000185`](https://nemar.org/dataset/nm000185) | [Sleep-EDF Database Expanded, Sleep Cassette study (PhysioNet sleep-edfx 1.0.0), loaded with braindecode SleepPhysionet (EEG channels Fpz-Cz / Pz-Oz only)](https://physionet.org/content/sleep-edfx/1.0.0/) | training | EEG |
+
+- Supervised AttnSleep (5 classes), channel Fpz-Cz only. SleepPhysionet subject_ids=[0..5], recording_ids=[1, 2] -> 12 recordings, but split ids are recording indices: train = indices 0-3 = subjects 0 and 1, both nights (SC4001, SC4002, SC4011, SC4012); valid = indices 4-5 = subject 2, both nights (SC4021, SC4022), used by EarlyStopping(monitor='valid_bal_acc', patience 20, load_best=True). Subjects 3, 4, 5 are loaded but never used. The tutorial text and the trainer's W&B config say 'subjects 0-3 train / 4-5 valid', which is not what the code does. Split is BaseConcatDataset.split(dict(train=..., valid=...)), whose values are RECORDING INDICES into the concatenated dataset, not subject ids (braindecode/datasets/base.py split: `[self.datasets[ds_ind] for ds_ind in ds_inds]`); SleepPhysionet orders recordings subject by subject, night 1 before night 2 (mne fetch_data loops `for subject in subjects`).
+- Sources: braindecode git 05e606dd^:scripts/train_and_push_tutorial_checkpoints.py (offline trainer that produced the tutorial checkpoints; removed from the repo in 05e606dd, 2026-04-19, and moved to https://gist.github.com/bru...; braindecode a283ae09:braindecode/datasets/base.py (BaseConcatDataset.split); W/bd/examples/applied_examples/plot_sleep_staging_eldele2021.py; braindecode check: https://huggingface.co/braindecode/plot_sleep_staging_eldele2021/resolve/main/history.json (summed batch sizes, epoch 1).
+