Pseudonymised human neurophysiological recordings, and the video clips the participants were shown — some of which are deliberately distressing. Every request is reviewed by HABS and access is granted for research use.

These are recordings of human participants, released in pseudonymised form. By requesting access you agree that you will not attempt to re-identify any participant, will not redistribute the recordings or any derivative from which a participant could be identified, will use them only for the research purpose you state below, and will delete your copy when that work is finished. The same applies to the stimulus clips: they are provided so that trials can be read against what was shown, and are not yours to republish. HABS may withdraw access at any time.

Multimodal EEG, physiology and eye-tracking during emotional video viewing

Synchronised brain and physiological recordings from 355 participants, collected while they watched videos chosen to elicit emotion, rested, and performed an artefact calibration task. Collected and released by HABS — Human Augmented Brain Systems.

DATASET.md is the guide. It describes the layout, the signals, the event timing, and every column of the three tables. Read it before you write anything.

What is in it


`CURATED/`	355 participants, unprocessed recordings — 178.07 GB
`PREPROCESSED/`	354 participants, cleaned and preprocessed — 173.07 GB
`videos/`	the 256 clips they were shown — 4.23 GB
`labels_train.csv`	29,514 video trials, with ratings and emotion labels
`markers.csv`	marker-code lookup
`participants_train.csv`	cohort, age and gender, one row per participant

videos/ is the stimulus set itself, so a trial can be read against what the participant was actually looking at. The rest is signal: each recording is a .bdf file with a matching .record_map sidecar describing its channels, units and sampling rates, and a per-day _metadata.json holding the session events. 64-channel EEG at 1,000 Hz, with ECG, skin conductance, respiration, gaze and pupil size. Keep every BDF beside its record map.

The clips they watched

videos/ holds all 256, one folder per theme, and videos/videos.csv joins them to the trials. The video column of labels_train.csv is the filename, the folder is not in that table — the folder names are French and video_theme is the English of the same thing, souffranceanimale and animal_suffering — so take the path from videos.csv rather than building it:

videos = pd.read_csv("videos/videos.csv").set_index("video")
labels = pd.read_csv("labels_train.csv", na_values=["TESTSUBJECT"])
labels["clip"] = labels["video"].map(videos["path"])

videos.csv also carries each clip's duration, resolution and the number of trials and participants that rest on it, so a clip shown four times is visible as such before anything is loaded.

Four clips in the folder appear in no trial: naked_4.mp4, sdf_4.mp4, souffranceanimale_4.mp4 and souffranceanimale_5.mp4. They are part of the stimulus set and were never shown in this cohort. One clip is renamed: the container spells it Injustice_2.mp4 and the table spells it injustice_2.mp4, and the table wins, because the table is what you read.

These are emotion-elicitation stimuli and some are meant to be unpleasant. Among the themes are injury and surgery, death and disaster, war, animal suffering, hate symbols, bullying and discrimination, and one of nudity. They were chosen to provoke a measurable response, not to be watched casually.

Download only what you need

It is 355 GB in total, so take a slice.

pip install huggingface_hub
hf auth login

# one participant
hf download HABSteam/multimodal-eeg-emotion --repo-type dataset --local-dir ./data \
    --include "CURATED/P001/*"

# every resting-state recording (~20 MB each)
hf download HABSteam/multimodal-eeg-emotion --repo-type dataset --local-dir ./data \
    --include "CURATED/*/*resting-state*"

# the tables, without any recordings
hf download HABSteam/multimodal-eeg-emotion --repo-type dataset --local-dir ./data \
    --include "*.csv"

# every stimulus clip (4.23 GB), or one theme's worth
hf download HABSteam/multimodal-eeg-emotion --repo-type dataset --local-dir ./data \
    --include "videos/*"

Resting-state recordings are about 20 MB, artefact about 45 MB, and emotion-video between 150 and 280 MB.

Held-out participants

Twenty-nine participants are held out for evaluation. Their recordings are not in this release, and in labels_train.csv and participants_train.csv every one of their fields reads TESTSUBJECT in place of a value — the trial rows are still there, so the participant, day, video and markers are visible, but no rating, no self-report and no emotion label is.

That sentinel is a string in otherwise numeric columns, which is enough to make rating_valence and four others load as text. Read the tables with it treated as missing:

labels = pd.read_csv("labels_train.csv", na_values=["TESTSUBJECT"])
labels = labels.dropna(subset=["final_emotion_label"])   # the 317 participants with labels

Before you report an accuracy

Recordings share videos, and repeated sessions or participant IDs introduce dependencies. A model evaluated on videos that also appear in training can recognise video-specific patterns, so its accuracy alone does not establish that it decodes emotion. Choose splits that match the question you are asking, and say which you chose.

final_emotion_label is derived from several annotation sources and may differ from a participant's own first-ranked report. It is not a direct measurement of subjective experience.

Terms

Pseudonymised human neurophysiological data, released under the data use agreement you accept when requesting access. Do not attempt to re-identify participants, do not redistribute the recordings, and delete your copy when your work is done. HABS may withdraw access at any time.

Controller: HABS — Human Augmented Brain Systems, SASU, 34 rue Laffitte, 75009 Paris, RCS Paris 953 261 583. See the HABS privacy policy. Data-protection questions and requests under the GDPR: dpo@habs.tech.

Downloads last month: 13,227

Total file size:

355 GB