EEG/EMG Foundation Challenge 2026
Preparation map · start kits + submissions

Your participant guide.

This guide takes you from choosing a track to submitting a trained model. NeuralBench offers optional start kits, public baselines, and a development environment, but you may use your own pipeline. Registration, submission rules, uploads, and official evaluation are handled on Codabench.

See the workflow ↓ Explore NeuralBench start kits ↗
Three interfaces · one workflow

Know where to go.

This website is your map. NeuralBench is the recommended optional preparation environment. Codabench is the official participation and evaluation platform.

Competition website

Understand the challenge.

Compare tracks, datasets, metrics, prizes, rules, dates, and organizers. Use this site to find the right next destination.

Explore the tracks →
NeuralBench · recommended path

Start from a reference.

Use the NeuralBench start kits to access public task examples, reproduce baselines, and develop or train models in a common environment.

Choose a NeuralBench start kit ↓
Codabench · official submission

Register and submit.

Read the Participation guidelines, register for each track, upload your trained model package, and receive official scores.

Open a Codabench track portal ↓
Recommended · optional

Warm up with the NeuralBench start kits.

Each NeuralBench guide is a track-specific start kit with public development data, explicit train, validation, and test splits, a baseline, and a reference result. Following it once gives you a verified pipeline before you introduce your own model, but it is not required. The reference results below belong to the named development datasets and are not Codabench warm-up leaderboard scores.

Track 01
Optional start kit

EEG-to-Image

Open Track 01 on NeuralBench ↗
Quick startSetup and test commands
Terminal
pip install neuralbenchneuralbench eeg image --downloadneuralbench eeg image --prepareneuralbench eeg image -m eegnet --debug

Install NeuralBench once, download the public data, prepare the cache, and verify the pipeline with a short EEGNet run.

Public-data referenceBaseline results

Baseline results obtained on THINGS-EEG2 (Gifford2022Large), using the split defined in the Track 01 NeuralBench start kit.

BaselineFamilyParamsTrainReference scoreConfiguration
Chancereference00 GPU-hTop-5 2.22 ± 0.31Not applicable
EEGNetLawhern et al., 2018CNN0.04 M2 GPU-hTop-5 28.13 ± 0.14configs/img/eegnet.yaml
REVEElouahidi et al., 2025Foundation14 M0.5 GPU-h (probe)Top-5 84.75 ± 0.38configs/img/reve_frozen.yaml

Reference values: Mean ± standard deviation from NeuralBench (Banville et al., 2026), Table 1.

Track 02
Optional start kit

BCI decoding

Open Track 02 on NeuralBench ↗
Quick startSetup and test commands
Terminal
pip install neuralbench 'moabb>=1.7.1'neuralbench eeg motor_imagery --downloadneuralbench eeg motor_imagery --prepareneuralbench eeg motor_imagery -m eegnet --debug

Install the MOABB data dependency, download the public analogue, prepare its cache, and verify the EEGNet pipeline.

Public-data referenceBaseline results

Baseline results obtained on Stieger 2021 (Stieger2021Continuous), a four-class motor-imagery dataset, using the split defined in the Track 02 NeuralBench start kit. Codabench currently uses Dreyer 2023 for warm-up evaluation, so these are not Codabench warm-up scores.

BaselineFamilyParamsTrainReference scoreConfiguration
Chancereference00 GPU-hBal. Acc 24.81 ± 1.03Not applicable
EEGNetCNN0.04 M4 GPU-hBal. Acc 58.58 ± 0.34configs/bci/eegnet.yaml
REVEFoundation14 M1 GPU-h (probe)Bal. Acc 68.04 ± 0.73configs/bci/reve.yaml

Reference values: Mean ± standard deviation from NeuralBench (Banville et al., 2026), Table 1.

Track 03
Optional start kit

Sleep onset

Open Track 03 on NeuralBench ↗
Quick startSetup and test commands
Terminal
pip install neuralbenchneuralbench eeg sleep_onset --downloadneuralbench eeg sleep_onset --prepareneuralbench eeg sleep_onset -m eegnet --debug

Install NeuralBench once, download Sleep-EDF, prepare the cache, and verify the sleep-onset pipeline with a short EEGNet run.

Public-data referenceBaseline results

Baseline results obtained on Sleep-EDF (Kemp2000Analysis), using the participant-level split defined in the Track 03 NeuralBench start kit.

BaselineFamilyParamsTrainReference scoreConfiguration
Chancereference00 GPU-hbMAE 205.42 ± 0.01 sNot applicable
EEGNet-sleepCNN0.04 M4 GPU-hbMAE 143.30 ± 0.40 sconfigs/sleep/eegnet.yaml
REVE-sleepFoundation14 M1 GPU-h (probe)bMAE 134.89 ± 2.02 sconfigs/sleep/reve.yaml

Reference values: Mean ± standard deviation from NeuralBench (Banville et al., 2026), Table 1. Here, bMAE is the unweighted mean of the MAE computed within each non-empty time-to-onset bin.

Track 04
Optional start kit

EMG-to-Pose

Open Track 04 on NeuralBench ↗
Quick startSetup and test commands
Terminal
pip install neuralbench 'eegdash>=0.8.2'neuralbench emg pose -m vemg2pose --downloadneuralbench emg pose -m vemg2pose --prepareneuralbench emg pose -m vemg2pose --debug

Install the EEGDash data dependency, then use VEMG2Pose for downloading, caching, and the debug run so every window has the required context.

Public-data referenceBaseline result

Baseline result obtained on public EMG2Pose (Salter2024Emg2pose), using the task split defined in the Track 04 NeuralBench start kit.

BaselineFamilyParamsTrainReference scoreConfiguration
NeuroPoseLiu et al., 2021Pose CNN-paperMAE 17.5 ± 1.5°tasks/emg/pose/config.yaml

Reference values: Mean ± standard deviation from NeuralBench (Banville et al., 2026), Table 1.

Open the NeuralBench challenge hub ↗ New to NeuralBench? Start here ↗
Your model · your pipeline

Train where you work best.

Both paths are valid. What matters is that you finish with a trained model that respects the task and can be exposed through the Codabench submission contract.

Recommended path

Build with NeuralBench.

Reuse its data access, task definitions, metrics, and reference baselines. You can reproduce a published model, add your own architecture, or build upon the provided workflow.

Explore NeuralBench ↗
Independent path

Use your own pipeline.

Train with your preferred codebase and permitted data. NeuralBench does not need to be part of your training stack or final submission, provided your trained model satisfies the track input, output, and packaging contract.

Review track requirements →

Same destination: Whatever the training path, Codabench receives a trained model for final evaluation.

From trained model to leaderboard

Package and submit on Codabench.

Once your model is trained, open the Codabench Participation tab for your track. It introduces you to the submission contract. Follow it to package your model, upload it through My Submissions, and confirm that your score reaches the leaderboard.

  1. 1Open ParticipationUse the dedicated tab for your chosen track below.
  2. 2Follow the contractFind the details on Codabench under Get Started → Participation, then package the trained model and required inference files.
  3. 3Upload the ZIPSubmit it through My Submissions during warm-up.
  4. 4Check the resultConfirm that evaluation succeeds and the score appears.

Two evaluation stages: Warm-up scores use public test data and are indicative. Codabench switches to the private 2026 cohorts for the sealed phase, and only sealed scores determine the final ranking.