PKUEEG directory index (2026-09-17)
==================================
Data acquisition institution: SHRC, PKU.

sub-XX/ses-dayY/eeg/                 Raw continuous BrainVision EEG and sidecars
participants.tsv / participants.json Participant metadata
stimuli/audio/                      Fifty shared story recordings
stimuli/transcripts/                Story text
stimuli/stimuli.tsv / stimuli.json   Shared stimulus index
derivatives/preproc_40hz/            1-40 Hz EEG, sampled at 250 Hz
derivatives/preproc_8hz/             1-8 Hz EEG, sampled at 128 Hz
derivatives/preprocessing_reports/  Availability of ICA/template processing records
derivatives/stimulus_features/      Envelope, Mel, wav2vec2 and BERT; static-vector status
derivatives/stimulus_annotations/   MFA, Pinyin and legacy word boundaries
code/                              Processing, feature extraction and structural/BIDS checks

Release scope: behavioral comprehension accuracy and technical-validation
analyses/results are deferred to a future dataset update. Historical/internal
validation reports are maintained outside this release. Anonymous participant
demographics and experimental event metadata remain included. The scope is also
recorded in code/config/release_scope.json.

Stories are allocated 17/16/17 across days (01-17, 18-33, 34-50).
Raw recordings remain continuous by day; stories are not separate raw runs.
Feature/annotation index paths are relative to the dataset root. Stimulus index
audio/text paths remain relative to stimuli/. No eye-tracking data or measured
individual electrode coordinates are implied.

The directory reorganization on 2026-09-16 moved existing files in place, without
copying the EEG or feature payloads or changing scientific values. A later,
metadata-only correction on the same date changed channels.tsv unit labels from
V to µV to match the BrainVision headers. It did not rescale any signal values.
No missing arrays were invented.
Earlier reports/checksums are preserved outside the public release directory.
Publication readiness is assessed separately; a directory scaffold is not evidence
of completeness. GitHub should host code/documentation, not this full data payload.

Original study description (paths updated)
-----------------------------------------
PKUEEG

PKUEEG contains anonymized EEG from 25 healthy right-handed native Mandarin
speakers during continuous speech perception and eyes-open rest. All participants
completed three recording sessions. Day 1 contains stories 1-17, Day 2 stories
18-33, and Day 3 stories 34-50. Each session includes one approximately 300-s
eyes-open resting interval.

Ethics
------
The protocol was approved by the Institutional Review Board of Peking University
(IRB00001052-25045). All participants provided written informed consent before
data collection. The EEG data were anonymized before release.

Acquisition
-----------
Days 1 and 2 were recorded at 1000 Hz with a NeuroScan SynAmps2 system using 62
scalp EEG channels and two EOG channels. Day 3 was recorded at 1000 Hz with a
NeuSen Wireless EEG/ERP system manufactured by Neuracle, with 59 scalp EEG
channels and five auxiliary channels. Recording system, session order, and story
set are confounded and should not be interpreted as independent experimental
factors. Standard montage labels are provided, but no participant-specific
electrode coordinates were collected.

Events
------
Each events.tsv contains one selected interval per story and one rest interval.
The original hardware marker stream remains in the BrainVision marker file. When
duplicate or accidental markers were present, the released interval used the last
adjacent marker pair whose separation was at least 150 s, matching the segmenting
rule used for the preprocessing derivatives.

Stimuli and behavior
--------------------
Fifty Mandarin story audio files and matching transcripts are included. The
audio was synthesized with the Guagua Audiobook platform. Stories 1-25 use the
Dapiaoliang female preset; stories 26-50 use the Huazai male preset.
Behavioral comprehension accuracy is planned for a future update. Story-level
stimulus derivatives include a 100-Hz speech envelope; uniform 50-Hz Mel, BERT
and Wav2Vec 2.0 representations; 12-layer word-level BERT vectors; and MFA
TextGrid, Pinyin-Hanzi and corrected word timing annotations. Detailed axes,
dimensions and provenance are recorded under derivatives/stimulus_features and
derivatives/stimulus_annotations.

Derivatives
-----------
The release includes NPZ derivatives at 1-40 Hz/250 Hz and 1-8 Hz/128 Hz. Signal
values are stored in microvolts. The Day-3 raw BrainVision data and 40-Hz NPZ
segments were corrected by a factor of 1e-6 because the source BDF physical unit
field was empty even though its physical range was expressed in microvolts. The
1-8-Hz derivative was regenerated uniformly from the corrected 40-Hz derivative.

Current contents and reproducibility limits
------------------------------------------
This copy contains legacy derivatives; it is not the later repair workflow's
output. The included EEG preprocessing entry point does not implement bad-segment
rejection. Extreme amplitudes remain in some NPZ segments, and per-session ICA
models, excluded-component lists and convergence records are not included.
Raw Day-3 labels and derivative labels differ in capitalization; channel matching
must account for this. The existing preprocessing code still has known
case-sensitive montage and bad-channel matching limitations.

The release contains 100-Hz envelope, 50-Hz Mel/wav2vec2/time-aligned BERT, and
12-layer word-level BERT arrays. word2vec_100hz is a status directory without
arrays; a complete 100-Hz Mel/wav2vec2/BERT feature set is not included.
The legacy word-level BERT script does not exactly reconstruct every tested
released array because its terminal overlapping-window rule differs. Complete
extraction pipelines for the other released features and an end-to-end
reproducible MFA workflow are not included.

code/model_downloads/ contains a status README, not pinned model-download
workflows. The dependency list is incomplete. Internal validation reports are
kept outside the release and are not certification of scientific reproducibility. These limitations
were documented, not repaired, by the metadata/documentation correction.
Acquisition reference and ground remain unavailable, and hardware filter settings
are not documented; no unknown acquisition parameters have been inferred.

Publishing and download links
-----------------------------
See code/docs/PUBLISHING.md and code/docs/GITHUB_README.md for the GitHub/OpenNeuro release plan. No public URL or DOI has been assigned yet.
Anonymous participant age and sex match the previously confirmed participant table; no names or initials are included.