README.md

metadata

pretty_name: OLIVE Physiological Dataset (EEG + Pupil)
license: cc-by-4.0
tags:
  - eeg
  - fixation-related-potentials
  - pupillometry
  - eye-tracking
  - xr
  - bci
  - attention
size_categories:
  - 10K<n<100K
configs:
  - config_name: default
    data_files:
      - split: train
        path: data/frp_dataset.parquet
  - config_name: ern
    data_files:
      - split: train
        path: data/ern.parquet

OLIVE Physiological Dataset (EEG + Pupil, FRP + ERN)

What this is

Per-fixation, fixation-related-potential (FRP) examples from the OLIVE Wingman / SpaceShooter user studies (US1 offline simulation, US2 live deployment, US3 silent target-switch), for the release cohort of 25 participants. Each example pairs a fixation-locked EEG epoch and pupil epoch with its target label, incoming-saccade metadata, per-fixation implicit-evidence probability, and condition and block information.

Fields (one row per fixation-locked epoch)

field	type	description
subject_id	int	participant id
user_study	str	us1 / us2 / us3
session	int	recording-session index within (subject, study); usually 0
eeg	float32[20, 230]	fixation-locked EEG epoch, see montage/window below
pupil	float32[2, 80]	fixation-locked pupil-diameter epoch (left/right), see window below
y	int (0/1)	ground-truth fixated-item label (1 = target)
p_target	float in [0,1] or NaN	per-fixation implicit-evidence target probability (in [0,1]); NaN where per-subject EEG-evidence parameters are unavailable
p_target_quality	float in [0.4,1.0] or NaN	per-subject decoder quality (AUC), or NaN
item_id	int	fixated item id within its block
run	int	1-indexed fixation-generating-block counter from the TFRecord (block_idx = run - 1, see caveat below)
fixation_duration	float seconds or NaN	fixation duration from long_gaze.jsonl; NaN unless with_saccades=True and the join matched
saccade_amplitude	float degrees or NaN	incoming-saccade amplitude; NaN unless with_saccades=True and derivable
saccade_angle	float radians or NaN	incoming-saccade direction (atan2(dy, dx))
saccade_peak_velocity	float deg/s or NaN	incoming-saccade peak velocity
saccade_mean_velocity	float deg/s or NaN	incoming-saccade mean velocity
saccade_dx, saccade_dy	float or NaN	gaze-forward-vector delta components across the incoming saccade
condition	str	primary, as-published condition label (IE / E); use this for any published-table-facing analysis
condition_ra	str or None	secondary RA-proposed overlay label; UNRESOLVED, does not match published tables, transparency/audit only
block_idx	int	0-indexed block within (subject, study, session); -1 if not derivable (see caveat)
difficulty	int	adaptive difficulty level for block_idx, from meta.jsonl; -1 if not found
task	str	visual_search (calibration blocks, difficulty==-1) or spaceshooter (gameplay blocks, difficulty>=0); a per-BLOCK label, not per-study

block_idx caveat: iter_epochs does not carry a literal block index field. run was verified empirically (against on-disk meta.jsonl/block-directory counts for multiple subjects across us1/us2/us3) to be a 1-indexed running block counter, so block_idx = run - 1 is used. This has not been verified for every subject/session; treat block_idx/difficulty as best-effort, not a guaranteed-exact join. task is derived from difficulty (difficulty == -1 => visual_search, else spaceshooter) and therefore inherits the same best-effort/fallback behavior: a failed metadata lookup falls back to difficulty = -1 => task = visual_search.

EEG montage and epoch windows

• EEG: 20-channel B-Alert X24 subset, standard 10-20 layout, sampled at 256 Hz. Epoch window is fixation-onset-locked [-0.1, 0.8] s → 230 samples/channel.
• Pupil: 2-channel (left/right) pupil diameter, sampled at 20 Hz. Epoch window is fixation-onset-locked [-1.0, 3.0] s → 80 samples/channel.
• Both windows match release/dataset/extract_epochs.py's EEG_N_T_DEFAULT / PUPIL_N_T_DEFAULT constants (reused, not redefined, here).

Condition legend

• condition (primary, as-published): IE = Implicit+Explicit (EEG + shot events), E = Explicit-only (shot events, no EEG). This is the label used throughout the paper's published tables and figures.
• condition_ra: a secondary v1 RA-proposed relabeling overlay (adds Oracle / Control labels for a handful of subjects) that is UNRESOLVED against the published cohort and does not match the paper's tables (see release/dataset/attach_metadata.py's module docstring and _RA_OVERRIDES). Included for transparency/auditing only; always prefer condition for anything that should agree with the paper.

p_target generation

p_target is the per-fixation implicit-evidence target probability produced by the default decoder in release/olive/decode.py (DefaultDecoder), which reads per-subject parameters under eeg_priors/. It is NaN for subjects/rows without available parameters. Replace the decoder with your own; see the repository README.

Coverage

Built from subjects=[4, 5, 12, 18, 20, 28, 29, 31, 33, 34, 35, 36, 37, 39, 40, 46, 47, 48, 49, 51, 52, 53, 54, 55, 59], studies=['us1', 'us2', 'us3'], with_saccades=False.

• Total examples: 58184
• Target-label rate (y==1): 0.4035 (23478 target / 34706 non-target)
• p_target valid (non-NaN): 18246 rows, mean=0.5059; NaN: 39938 rows
• p_target_quality (valid rows): mean=0.7669, range=[0.5375, 0.8851]
• Subjects with zero examples across all studies: 0 ([])
• Subjects with >0 examples in at least one study: 25 ([4, 5, 12, 18, 20, 28, 29, 31, 33, 34, 35, 36, 37, 39, 40, 46, 47, 48, 49, 51, 52, 53, 54, 55, 59])

Full per-subject x per-study counts: see coverage.csv (same directory as this card). Table (subject_id x study, total = row sum):

subject_id	us1	us2	us3	total
4	725	925	1046	2696
5	591	0	0	591
12	564	0	863	1427
18	796	745	0	1541
20	1071	1407	994	3472
28	840	867	863	2570
29	633	901	1025	2559
31	520	0	0	520
33	662	598	560	1820
34	997	1109	1305	3411
35	778	1023	1217	3018
36	1022	1332	1466	3820
37	1014	932	1026	2972
39	1226	1755	1609	4590
40	875	693	795	2363
46	783	0	1217	2000
47	612	0	0	612
48	1412	2049	1593	5054
49	759	1961	1488	4208
51	1049	1332	798	3179
52	607	0	0	607
53	1013	1473	0	2486
54	946	0	0	946
55	924	0	0	924
59	798	0	0	798

Saccade fields caveat (with_saccades)

saccade_* and fixation_duration are all-NaN by default (with_saccades=False). Deriving them requires loading a session's raw .p eye-tracking recording (up to ~1-3GB) and joining each epoch's TFRecord-clock fix_time_s against long_gaze.jsonl's LSL-clock t_onset by nearest-match within item_id, subject to a tolerance (default 1.0s); these two clocks are not guaranteed to share an origin, so even with with_saccades=True, a session whose clocks do not align will legitimately yield all-NaN saccade fields for every epoch in that session; this is a documented limitation of the join, not a bug. A second caveat: for the small number of subjects with multiple numbered "session" subdirectories under one study directory but only one physical .p file at the study-directory root (observed for subject 5 / us1), the join falls back to that one shared .p for every session under that study directory, which can pool fixation events across sessions. See release/dataset/export_hf.py's module docstring for full detail.

Consent

All participants in the published cohort provided informed consent under the study's IRB protocol for their de-identified physiological (EEG, pupil), behavioral, and gaze data to be included in a public research dataset release. No directly identifying information (name, contact info, raw video) is included in this export; subject_id is a study-internal integer id, not a real-world identifier.

License

TBD: placeholder, CC-BY-4.0 (pending final confirmation from the study PI / IRB before public release).

Citation

Placeholder: update with the final paper citation before public release:

@inproceedings{olive-physio-2026,
  title     = {OLIVE: [paper title TBD]},
  author    = {[authors TBD]},
  booktitle = {[venue TBD]},
  year      = {2026},
}

ERN variant

A second HuggingFace config (ern), loadable via load_dataset("ApocalyVec/olive-physio", "ern"), of per-shot response-locked ERN (error-related negativity) epochs, distinct from the fixation-locked FRP epochs in the default config above.

What this is

One row per in-game shot event (enemy hit = correct, friendly-fire hit = error) from the OLIVE Wingman / SpaceShooter user studies (US1, US2, US3), for the release cohort of 25 participants. Each row is a single-trial, response-locked EEG epoch around the shot event, labeled correct/error.

Fields (one row per shot event)

field	type	description
subject_id	int	participant id
study	str	us1 / us2 / us3
session	str	recording-session .p file stem
eeg	float32[20, 205]	response-locked EEG epoch, uV, see window below
label	int (0/1)	0 = correct (enemy hit), 1 = error (friendly-fire hit)
shot_time	float	LSL timestamp of the shot event
montage	list[str]	B-Alert channel names, in eeg row order
condition	str	primary, as-published condition label (IE / E), same as the default config's condition field; use this for any published-table-facing analysis

Epoch window and filtering

• EEG: 20-channel B-Alert X24 subset (same montage as the FRP config), sampled at 256 Hz.
• Window: response-locked (shot-event-locked) [-200, 600] ms → 205 samples/channel.
• Filter: continuous zero-phase 4th-order Butterworth bandpass, 0.5-30.0 Hz, applied to the full continuous recording before epoching (per-epoch filtering of an ~800 ms window is invalid for a 0.5 Hz high-pass, which needs several seconds of settling).
• Baseline window: [-200, 0] ms (pre-response) is the conventional ERN baseline period included in the epoch; the exported eeg array is the filtered epoch as-is and is not baseline-corrected; apply baseline correction (subtract the [-200, 0] ms mean per channel) yourself if your analysis requires it.
• Label convention: shot events come from Unity.ReNa.EventMarkers row 2 (DTN-coded); DTN==1 (friendly fire) → label=1 (error), DTN==2 (enemy) → label=0 (correct). Verified identical across US1/US2/US3 (see release/dataset/ern/extract_ern.py's module docstring for the per-subject verification counts).

Availability

Available for all three studies: US1 (offline simulation), US2 (live deployment), US3 (silent target-switch). Coverage varies by subject/study; some subjects have zero epochs in a given study (no session recorded, or no .p file with usable shot events); see the coverage table below and ern_coverage.csv (same directory as this card) for exact per-subject counts.

Coverage

Built from subjects=[4, 5, 12, 18, 20, 28, 29, 31, 33, 34, 35, 36, 37, 39, 40, 46, 47, 48, 49, 51, 52, 53, 54, 55, 59], studies=['us1', 'us2', 'us3'].

• Total examples: 29178
• Error rate (label==1): 0.2352 (6864 error / 22314 correct)
• Subjects with zero examples across all studies: 1 ([47])
• Subjects with >0 examples in at least one study: 24 ([4, 5, 12, 18, 20, 28, 29, 31, 33, 34, 35, 36, 37, 39, 40, 46, 48, 49, 51, 52, 53, 54, 55, 59])

Full per-subject x per-study correct/error counts: see ern_coverage.csv. Table (subject_id x study, total = row sum):

subject_id	us1_correct	us1_error	us2_correct	us2_error	us3_correct	us3_error	total
4	218	100	404	141	483	118	1464
5	224	120	0	0	0	0	344
12	158	80	0	0	0	0	238
18	190	85	327	199	0	0	801
20	162	81	388	190	412	117	1350
28	258	103	416	149	502	137	1565
29	291	143	544	215	651	148	1992
31	300	137	0	0	0	0	437
33	246	109	613	123	513	106	1710
34	362	89	683	111	573	115	1933
35	300	113	377	57	226	31	1104
36	427	95	697	120	687	85	2111
37	381	89	714	152	684	109	2129
39	273	123	651	232	566	210	2055
40	264	66	0	0	540	105	975
46	316	134	0	0	566	213	1229
47	0	0	0	0	0	0	0
48	305	109	627	228	605	165	2039
49	92	43	563	247	573	193	1711
51	229	84	438	142	392	140	1425
52	182	65	0	0	0	0	247
53	279	81	676	198	0	0	1234
54	313	114	0	0	0	0	427
55	240	119	0	0	0	0	359
59	213	86	0	0	0	0	299