Rules & FAQ.
Seven binding rules, followed by the questions participants ask most.
Eligibility, tracks, submission caps, external data, compute, the reproducibility audit, and authorship are binding. Every participant is expected to have read the rules before submitting.
Seven rules. Read them before you submit.
Every rule is sourced from the competition proposal. Submitting means you have read and accepted them.
If this page contradicts something you read elsewhere on the site, this page wins — and please tell us, by email or Discord.
-
Eligibility
Anyone can enter. Organisers cannot win prizes.
Industry, academia, students, independents — all eligible, and cross-institution teams are encouraged. Organisers and their direct lab members may submit, but their entries appear as
Organizers · referenceon the leaderboard and are ineligible for cash prizes. -
Tracks
One team may enter all four tracks.
Each track accepts independent submissions per team and is scored on its own leaderboard, under the same harness and the same audit.
-
Submission cap
Five submissions a day in warm-up, two in the sealed final.
Warm-up (Sep 16 – Oct 15, 2026): 5 submissions per team per day. Sealed final (Oct 16 – Nov 16, 2026, AoE): 2 per team per day, so the prize roster reflects deliberate iteration rather than lottery search. The final NeurIPS ranking takes the best of your last five sealed submissions.
-
External data
Pre-train on any public dataset. Declare all of it.
Training data is suggested to stay within the track, but pre-training on any publicly available, redistributable dataset is allowed. The sealed test split is never allowed; closed clinical datasets are not allowed. Declare every external corpus, and a compute estimate, in the method description that ships with your final submission — the audit cross-checks it against a re-run of
neuralbench eeg <task>from your committed config.OKOpenNeuro · MOABB · public BIDSNot OKSealed test · private clinical -
Compute
Training compute is uncapped. Inference is 60 minutes on one GPU.
Train on whatever you have. The single constraint is inference: the scoring container must complete a full test pass in under 60 minutes on one H100 or H200 instance, so that audit cost stays bounded and per-team runtime stays comparable. AWS, the compute partner, provides complimentary instance hours to finalists.
Inference budget60 min · H100 / H200TrainingNo capAudit hardwareCodabench container -
Reproducibility audit
The top three per track must replay within ±2σ.
After the sealed final (date announced after the summer), the top-3 of every track go through a reproducibility audit: we re-run your training pipeline from the committed config and re-score the resulting weights against the sealed split. Scores within ±2σ of your submitted number stay on the prize roster; scores outside the tolerance drop off the prize roster but remain on the public board for context. Arnaud Delorme (EEGLAB) chairs the audit.
Tolerance±2σ on metricAudit leadArnaud Delorme (EEGLAB)WindowAnnounced after summer -
Authorship
Reproducible top teams are named authors on the competition report.
Top-ranked teams that pass the audit and submit a method description, training and inference code, and pre-training disclosures are invited as named authors on the PMLR competition report. Consortium authorship is offered to teams with complete, reproducible artifacts.
Questions people ask.
Answers, not rules: nothing here overrides the seven rules above.
Schedule shifts and analysis-group merges are announced on the website and Discord before they take effect.
DatasetsWhen is the Track 3 Muse Sleep-Onset dataset available, and how do I access it?
The four NEW 2026 competition releases - Muse Sleep-Onset (Track 3), Graz/BrainHero (Track 2), and the Alljoined and EMG2Pose evaluation splits - become available when the competition opens on September 16, 2026. From that date, neuralbench eeg sleep_onset --download pulls the official Muse cohort directly. There is no separate platform, access link, or manual permission step beyond registering your team - just update your NeuralBench install after Sep 16.
Before September, neuralbench eeg sleep_onset --download returns the seed datasets (Sleep-EDF, PhysioNet Challenge 2018, HMC Sleep). These are starter-kit / analog data for building and validating your pipeline — development stand-ins, not the sealed evaluation set or the official Track 3 cohort. Release notices go to the website and Discord, so watch there for the exact drop.
Low participationWhat happens if a track has too few participants?
Low-participation tracks fold into the closest active analysis group for the post-competition report, but prizes are still awarded on that track's own leaderboard. Each track is scored independently, so low participation on one track never shrinks the field on another.
DelaysWhat if there's a delay?
Technical, data, or infrastructure delays shift all downstream deadlines by the same amount on the website and Discord. The reproducibility audit gate is the only milestone that can extend independently if a top submission needs more re-run time.
Data releaseWill the competition data be released?
After the challenge, new data is made publicly available in full or in part where consent, licensing, and provider policy allow. The hidden evaluation labels remain confidential indefinitely so the benchmark stays usable for future research.
Reach the organising team.
Email or Discord. Same response time either way.