Add files using upload-large-folder tool

README.md

@@ -1,67 +1,71 @@
 ---
-license: unknown
+license: cc0-1.0
 tags:
   - eeg
   - sleep-staging
   - in-ear-eeg
+  - scalp-eeg
+  - hdf5
 ---
 
 # EESM23-Processed
 
-Preprocessed HDF5 export of the **EESM23** (Aarhus in-ear EEG sleep) BIDS
-dataset — 10 subjects, 2 nights each (`ses-001`, `ses-002`), one 30-second
-AASM scoring epoch per row (`Wake` / `N1` / `N2` / `N3` / `REM`; `Artefact`
-epochs dropped). `sub-006/ses-002` PSG is skipped — the source `.set` file is
-truncated on disk.
+Processed HDF5 export of the [EESM23 OpenNeuro dataset](https://doi.org/10.18112/openneuro.ds005178.v1.0.0). It contains paired in-ear EEG and scalp EEG sleep-staging samples from the scored `ses-001` and `ses-002` recordings of 10 subjects.
 
 ## Preprocessing
 
-**Only** a 0.1–100 Hz band-pass + 50 Hz notch filter is applied, on the full
-continuous recording before slicing into 30 s epochs (to avoid per-epoch
-filter edge effects). Nothing else: **no re-referencing, no resampling, no
-channel renaming**. Channel names are kept exactly as in the source BIDS
-`channels.tsv`.
+Generated with `Ear-EEG-FM-Benchmark/dataset/preprocess_eesm23.py` using schema/eegfm version `0.4.0`:
 
-Device data-loss gaps (NaN samples) are linearly interpolated before
-filtering (a long FIR kernel otherwise smears each NaN across a wide window)
-and the true NaN positions are restored afterward, so `nan_fraction` still
-reflects genuine data quality rather than a filtering artifact.
+- 0.1–100 Hz band-pass and 50 Hz notch filtering on each continuous recording
+- no re-referencing, resampling, or channel renaming
+- each labeled 30-second AASM scoring event is split into seven non-overlapping 4-second windows; the final 2 seconds are unused
+- classes: `Wake`, `N1`, `N2`, `N3`, `REM`; `Artefact` events are dropped
+- windows containing non-finite samples are dropped
+- all signal values are stored as `float32` microvolts at 250 Hz
 
-Generated by `dataset/preprocess_eesm23.py` in the EEGFM repo.
+Brief device data-loss gaps are interpolated before filtering to prevent FIR-kernel contamination, after which the original NaN positions are restored before window selection.
 
 ## Files
 
-| file | channels | epochs (N) | sfreq | epoch length |
-|---|---|---|---|---|
-| `eesm23-in-ear-eeg.h5` | 4: RB, RT, LB, LT (acq=earEEG, original names) | 16553 | 250 Hz | 30.0 s (7500 samples) |
-| `eesm23-scalp-eeg.h5` | 8: M1, F3, C3, O1, M2, F4, C4, O2 (acq=PSG) | 15526 | 250 Hz | 30.0 s (7500 samples) |
+| File | Channels | Shape `(N, C, T)` | Size |
+|---|---|---:|---:|
+| `eesm23-in-ear-eeg.h5` | RB, RT, LB, LT | `(111594, 4, 1000)` | 1.70 GiB |
+| `eesm23-scalp-eeg.h5` | M1, F3, C3, O1, M2, F4, C4, O2 | `(95339, 8, 1000)` | 2.88 GiB |
 
 Label distribution:
 
-| | Wake | N1 | N2 | N3 | REM |
-|---|---|---|---|---|---|
-| in-ear | 1501 | 1369 | 7792 | 2667 | 3224 |
-| scalp | 1375 | 1293 | 7329 | 2494 | 3035 |
+| Modality | Wake | N1 | N2 | N3 | REM |
+|---|---:|---:|---:|---:|---:|
+| in-ear | 9,725 | 9,268 | 52,746 | 17,983 | 21,872 |
+| scalp | 8,203 | 7,898 | 45,010 | 15,645 | 18,583 |
 
-## HDF5 schema (v0.2)
+Source-data notes:
 
-```
-/data          (N, C, T)   float32   signal, µV
-/durations     (N,)        int64     valid samples per epoch (== T here, fixed-length)
-/nan_fraction  (N,)        float32   fraction of non-finite samples in the epoch
-/labels        (N,)        int64     index into attrs['class_names']
-/subject       (N,)        str       'sub-001' ...
-/session       (N,)        str       'ses-001' / 'ses-002'
-/task          (N,)        str       'sleep'
-/run           (N,)        str       '' (unused, sessions are not run-qualified)
-/trial_id      (N,)        int64     row index in the source scoring events.tsv
-/ch_names      (C,)        str       channel names, as in source BIDS channels.tsv
-
-attrs: sfreq, n_class, class_names, unit ('uV'), eegfm_version,
-       source_bids_path, bids_dataset_name, preprocess_config_json, created_at
-```
+- `sub-006/ses-002` PSG is skipped because the source `.set` file is truncated and cannot be read.
+- `sub-002/ses-001` and `sub-007/ses-002` PSG yield no windows after the non-finite-sample quality rule.
+
+## HDF5 schema (v0.4)
 
-`preprocess_config_json` (per file) records the exact filter settings, e.g.:
-```json
-{"acq": "earEEG", "sessions": ["001", "002"], "epoch_sec": 30.0, "filter_low": 0.1, "filter_high": 100.0, "notch": 50.0}
+```text
+/data                 (N, C, 1000) float32
+/durations            (N,)         int64
+/nan_fraction         (N,)         float32
+/labels               (N,)         int64
+/sample_id            (N,)         int64
+/subject              (N,)         string
+/session              (N,)         string
+/task                 (N,)         string
+/acquisition          (N,)         string
+/run                   (N,)         string
+/recording_id          (N,)         string
+/trial_id              (N,)         int64
+/event_id              (N,)         int64
+/split_group_id        (N,)         int64
+/window_start_sample   (N,)         int64
+/window_stop_sample    (N,)         int64
+/ch_names              (C,)         string
 ```
+
+`event_id` and `split_group_id` identify the source 30-second scoring row. Keep all windows sharing a split group together when constructing train/validation/test partitions.
+
+Important attributes include `sfreq`, `class_names`, `unit`, `eegfm_version`, `preprocess_config_json`, `split_group_kind`, and `window_reference`.

eesm23-in-ear-eeg.h5

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:2696190f5ffdf78664078586eef1f1ab4d013d4e424e76b4aa3812fe85a7720b
-size 1990377120
+oid sha256:9fc0b2b7d97a3f6248a1cccea67043095f8e98ae4388b60097b527e45890f193
+size 1828928328

eesm23-scalp-eeg.h5

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:83d6cc80a09fd79bab63968970e059383deb1d193e8d0f0ddc7174931e7b411b
-size 3730016416
+oid sha256:34de1fc13780a9de42cb1079137785a686c6f997db2c1dc64e1f927c33076396
+size 3088115976