# Ear-EEG FM Benchmark — preprocessed datasets

This repository hosts preprocessed HDF5 files for the Ear-EEG FM Benchmark
(https://github.com/zhikaili1150/Ear-EEG-FM-Benchmark).

Each `.h5` file follows the schema:
```
data       (N, C, T) float32    µV
labels     (N,)      int64       0..n_class-1
subject    (N,)      object      'sub-NN'
session    (N,)      object      'ses-NN'
trial_id   (N,)      int64
ch_names   (C,)      object      standard 10-20 names
attrs      sfreq, n_class, class_names, unit, preprocess_config_json, ...
```

Naming convention: `<dataset>-eeg.h5` = scalp variant, `<dataset>-ear-eeg.h5` = ear variant.

| file | task | trials | shape | sfreq | source |
|---|---|---|---|---|---|
| `saad-eeg.h5`       | binary AAD (left vs right)            | 13,500 | (N, 27, 4000) | 1000 Hz | https://zenodo.org/records/16536441 |
| `saad-ear-eeg.h5`   | binary AAD (cEEGrid around-ear)       |        | (N, 14, 4000) | 1000 Hz | https://zenodo.org/records/16536441 |
| `eesm23-eeg.h5`     | 5-class sleep (scalp PSG)             |        | (N,  8, 1000) |  250 Hz | https://openneuro.org/datasets/ds005178 |
| `eesm23-ear-eeg.h5` | 5-class sleep (cEEGrid ear-EEG)       | 56,967 | (N,  4, 1000) |  250 Hz | https://openneuro.org/datasets/ds005178 |
| `aarhus-aad-eeg.h5`     | binary AAD (left vs right, scalp)     |    960 | (N, 30, 42781) | 1000 Hz | https://zenodo.org/records/8082867 |
| `aarhus-aad-in-ear-eeg.h5` | binary AAD (left vs right, in-ear) |    960 | (N, 12, 42781) | 1000 Hz | https://zenodo.org/records/8082867 |

Use with the benchmark pipeline:
```bash
git clone git@github.com:zhikaili1150/Ear-EEG-FM-Benchmark.git
cd Ear-EEG-FM-Benchmark
mkdir -p data
huggingface-cli download <this repo> --repo-type dataset --local-dir data
DATASET=saad-eeg bash scripts/run_benchmark.sh
```