Instruction tuning datasets
This chapter covers the 16 datasets reported in the paper, spanning motor imagery, emotion recognition, SSVEP, covert speech, and mental workload. Dataset keys and label ordering follow configs/tasks.yaml.
Motor imagery
All registered motor-imagery datasets are converted to 200 Hz, mapped to the 65-channel montage, and stored as subject-keyed HDF5 files.
MI_OpenBMI
- Source: MATLAB files under
MI_OpenBMI/DB_mat/session1|session2/<subject>/EEG_MI.mat. - Subjects: 54 (
s01-s54); first 42 train/validation, last 12 test. - Labels: stored
0=Right,1=Leftaccording toconfigs/tasks.yaml; source labels are reduced by one. - Trials: concatenate train and test arrays from both acquisition sessions.
- Signal processing: construct epochs at 1,000 Hz, resample to 200 Hz, band-pass 0.3-40 Hz, then apply the shared clipping/scaling/montage pipeline.
- Output:
MI_OpenBMI.h5, grouped by subject.
MI_BCIC_IV2a
- Source: each subject has evaluation and training MATLAB files,
<subject>E.matand<subject>T.mat, underMI_BCIC_IV2a_mat/. - Subjects: 9 (
A01-A09); first 7 train/validation, last 2 test. - Labels: Left, Right, Foot, Tongue; MATLAB labels are reduced by one.
- Trial extraction: skip the first three MATLAB data blocks. Use event boundaries to extract 22 EEG channels, subtract the across-channel mean at each time point, and initially retain 0-6 seconds at 250 Hz.
- Signal processing: band-pass 0.3-40 Hz, resample to 200 Hz, retain seconds 2-6 (800 samples), then apply the shared pipeline.
- Output:
MI_BCIC_IV2a.h5.
MI_BCIC_Upperlimb
- Source:
Training set/sampleXX.matandValidation set/sampleXX.mat. - Subjects: 15; first 11 train/validation, last 4 test.
- Labels: Cylin, Sphe, Lumbrical from the one-hot source labels.
- Trial extraction: concatenate the published training and validation arrays. The code removes the final time sample before creating MNE epochs.
- Signal processing: resample from the file's
fsvalue to 200 Hz; band-pass 0.1-70 Hz; retain the final motor-imagery stage beginning at second 6, which is documented in code as 4 seconds; apply the shared pipeline. - Output:
MI_BCIC_Upperlimb.h5.
MI_ShanghaiU
- Source: five files
d1.mat-d5.matper subject. - Subjects: 25 (
S01-S25); first 20 train/validation, last 5 test. - Labels:
configs/tasks.yamlinterprets0=Right,1=Leftafter subtracting one from source labels. - Trial extraction: concatenate all five sessions.
- Signal processing: create 250 Hz epochs, drop
A1andA2, resample to 200 Hz, band-pass 0.3-40 Hz, and apply the shared pipeline. - Output:
MI_ShanghaiU.h5.
MI_HighGamma
- Source: one exported
.npzper subject containingx_data,y_data, sampling rate, and metadata. - Subjects: 14; first 10 train/validation, last 4 test.
- Builder labels: keep source classes 1-3 and convert to
0=Left,1=Right,2=Feet; the fourth source class is removed. - Training labels:
load_datasetremoves label 2 again, so the registered task is binary Left/Right. - Signal processing: drop mastoid channels
M1andM2, resample to 200 Hz, band-pass 0.3-40 Hz, and apply the shared pipeline. - Output:
MI_HighGamma.h5.
MI_Cho2017
- Source: exported
.npzfiles for 49 available subjects;s32,s46, ands49are excluded. - Split: first 40 subjects train/validation, final 9 test.
- Labels: Left and Right; labels are used without an offset.
- Signal processing: divide source values by 100, resample to 200 Hz, band-pass 0.3-40 Hz, append 200 samples by edge padding, and apply the shared pipeline.
- Output:
MI_Cho2017.h5.
MI_Shin2017A
- Source:
cnt.matandmrk.matfrom thewith occular artifactdirectory for each subject. - Subjects: 28;
S05is unavailable. The first 22 available subjects are train/validation and the last 6 are test. - Sessions: process source blocks with indices 0, 2, and 4.
- Signal processing: load at 200 Hz; remove
VEOGandHEOG; band-pass 0.3-40 Hz; notch at 50 Hz; epoch 0-4 seconds; then discard the first sample so the inclusive MNE epoch becomes 800 samples. Apply the shared pipeline. - Labels: marker codes are reduced by one to Left/Right.
- Output:
MI_Shin2017A.h5.
MI_PhysioNet
- Source: EDF runs
R03throughR14for 109 PhysioNet EEG Motor Movement/Imagery subjects. - Split: first 80 subjects train/validation, last 29 test.
- Signal processing: interpolate channels marked bad by the EDF reader, resample to 200 Hz, notch at 60 Hz, band-pass 0.3-40 Hz, strip trailing periods from channel names, and epoch annotations from 0 to 4 seconds. The code removes the trailing inclusive sample to produce 800 samples, converts signals to microvolts, and applies the shared pipeline.
- Label filtering: keep annotation IDs 2 and 3 and convert them to labels 0 and 1.
- Output:
MI_PhysioNet.h5.
Emotion
The registered emotion suite contains FACED and four SEED variants. FACED uses 10-second windows and a held-out-subject test split. SEED datasets use 4-second windows and predefined trial splits.
EMO_FACED
- Source:
Processed_data/subXXX.pklfor subjects000-122. - Channels: the two reference channels
A1andA2are dropped from the 32-channel source list. - Signal processing: create 250 Hz epochs, band-pass 0.1-70 Hz, resample to 200 Hz, and divide each recording into non-overlapping 10-second windows. Incomplete final windows are discarded.
- Labels: the 28 source trials are assigned to Anger, Fear, Disgust, Sad, Amusement, Inspiration, Joy, Tenderness, and Neutral using the fixed label array in the builder. Each trial label is repeated for all of its windows.
- Normalization: apply the shared pipeline per subject after segmentation.
- Split: subjects 000-099 form the train/validation pool; subjects 100-122 are test.
EMO_SEED_3_seg4
- Source: three MATLAB sessions for each of 15 subjects under
EMO_SEED_3/eeg_raw_data/. - Channels/rate: 62 channels at 200 Hz.
- Trial construction: for each of 15 trial IDs, concatenate that trial from all three sessions along time before filtering and segmentation.
- Signal processing: band-pass 0.1-70 Hz, notch at 50 Hz with width 2 Hz, read in microvolts, and divide into non-overlapping 4-second windows.
- Split by trial ID: trials 1-9 train, 10-12 validation, 13-15 test.
- Normalization: shared pipeline is run separately for each subject and split. After subjects are concatenated, values are clipped again to
[-10, 10]. - Labels: the builder converts raw
-1, 0, +1to stored0, 1, 2, giving Negative, Neutral, Positive. This order now matches bothEMO_SEED_3.py::TEXT_LABELSandconfigs/tasks.yaml.
EMO_SEED_4_seg4
- Source: three session directories, each containing one MATLAB file per each of 15 subjects.
- Channels/rate: 62 channels at 200 Hz.
- Signal processing: band-pass 0.1-70 Hz, notch at 50 Hz with width 2 Hz, and split each trial into non-overlapping 4-second windows.
- Split within each session: trials 1-16 train, 17-20 validation, 21-24 test.
- Labels:
0=Neutral,1=Sad,2=Fear,3=Happy, matchingconfigs/tasks.yaml. - Normalization: shared pipeline per subject and split; final arrays clipped to
[-10, 10].
EMO_SEED_5_seg4
- Source: three CNT sessions for each of 16 subjects under
EMO_SEED_5/EEG_raw/. - Channels: drop
VEO,HEO,M1, andM2, leaving the declared 62 channels. - Trial extraction: fixed start/end timestamps define 15 video trials in each session.
- Signal processing: resample to 200 Hz, band-pass 0.1-70 Hz, notch at 50 Hz with width 2 Hz, read microvolts, and split into non-overlapping 4-second windows.
- Split within each session: trials 1-5 train, 6-10 validation, 11-15 test.
- Labels:
0=Disgust,1=Fear,2=Sad,3=Neutral,4=Happy. - Normalization: shared pipeline per subject and split; final arrays clipped to
[-10, 10].
EMO_SEED_7_seg4
- Source: four CNT sessions for each of 20 subjects.
- Channels: drop
M1,M2,ECG,HEO, andVEO, leaving 62 channels. - Events: use annotation pairs as trial start/end positions. Two known recordings (
14_20221015_1.cntand9_20221111_3.cnt) reconstruct event times from CSV trigger logs. - Signal processing: resample to 200 Hz, band-pass 0.1-70 Hz, notch at 50 Hz with width 2 Hz, read microvolts, and split each of 20 trials per session into non-overlapping 4-second windows.
- Split within each session: trials 1-10 train, 11-15 validation, 16-20 test.
- Labels:
0=Happy,1=Surprise,2=Neutral,3=Sad,4=Disgust,5=Fear,6=Angerfrom the fixed video order. - Normalization: shared pipeline per subject and split; final arrays clipped to
[-10, 10].
Predefined HDF5 split
All four SEED task files are loaded directly from trainX/trainY, validX/validY, and testX/testY.
SSVEP, covert speech, and workload
These three datasets are also registered in configs/tasks.yaml and included in the standard instruction-tuning loop.
SSVEP_OpenBMI
- Source: the OpenBMI MATLAB structure under
MI_OpenBMI/DB_mat/session1|session2/<subject>/EEG_SSVEP.mat. - Subjects: 54; first 42 train/validation, last 12 test.
- Trials: concatenate the published train/test arrays from both sessions.
- Labels: source labels are reduced by one and interpreted as
0=12.0,1=8.6,2=6.6,3=5.4Hz. - Signal processing: create epochs at 1,000 Hz, resample to 200 Hz, band-pass 0.3-40 Hz, and apply the shared pipeline. The general 0.1-70 Hz filter is commented out; the current implementation uses the motor-imagery band.
- Output:
SSVEP_OpenBMI.h5. - Split: the first 42 subjects form the train/validation pool; the last 12 subjects are test.
CS_BCIC_Speech
- Source:
Training set/Data_SampleXX.matandValidation set/Data_SampleXX.matfor 15 subjects. - Labels: hello, help-me, stop, thank-you, yes, obtained by
argmaxover source one-hot labels. - Trial construction: transpose both source arrays to
(trial, channel, time), concatenate them, and append five time samples using edge padding. - Signal processing: apply the shared clipping, robust-scaling, and montage-mapping pipeline to the published epochs.
- Output:
CS_BCIC_Speech.h5, grouped by subject. - Split: for every subject, trials 0-249 are training, 250-299 validation, and 300 onward test; the subject and trial axes are then flattened.
Workload
- Source: two EDF files per subject,
SubjectXX_1.edfandSubjectXX_2.edf, for 36 subjects. - Labels: recording 1 becomes
0=Resting; recording 2 becomes1=Workload. - Channels: remove
ECG ECGandEEG A2-A1; remove the first four characters from all remaining source channel names. - Signal processing: resample to 200 Hz, band-pass 0.1-70 Hz, divide each continuous recording into non-overlapping 4-second windows, concatenate both conditions, and apply the shared pipeline.
- Output:
Workload.h5. - Split: subjects 00-31 form the train/validation pool; subjects 32-35 are test.