EEG/EMG Foundation Challenge · NeurIPS 2026
NeuralBench · track guides + baselines

Get prepared with NeuralBench.

NeuralBench is the preparation path for every track. To get prepared, install the framework, follow the track guide, reproduce a public baseline, and compare your result with the reference scores. The same platform supports the competition workflow once submissions open.

Open the NeuralBench challenge hub ↗
One-time setup

Install and configure.

Start here even if you plan to develop your model in a separate codebase. NeuralBench provides the common data loading, task definitions, metrics, and evaluation interface used by the challenge. Install it once, then configure stable locations for datasets, preprocessing caches, and results.

Install from PyPI pip install neuralbench

Use Python 3.12 or newer, preferably in a clean virtual environment. The base install is enough to download data, run the public tasks, load model weights, and save scores.

Installation and configurationPython, storage paths, and local or SLURM execution ↗ NeuralBench quickstartRun a first benchmark task end to end ↗
One guide per track

Open your track guide.

Choose a track, then follow its guide once without changing the model or configuration. Each guide explains the scientific task, the public starter data, the competition shift, and the exact commands needed to reproduce its baseline. This gives you a known working pipeline before you introduce your own model.

Public-data warm-up: Track 2 currently uses NeuralBench’s motor-imagery analog, and Track 3 uses Sleep-EDF until the competition releases are available. These runs validate your pipeline. Hidden competition scores will differ.

Same sequence for every track

Reproduce the baseline.

Copy the base command from your track guide and run the four actions in order. Complete the small debug run before full training. If the final score is close to the published reference, your data, model, and metric path are ready for experimentation.

  1. 1Download dataFetch the public starter dataset selected by the guide.--download
  2. 2Build the cacheApply the task preprocessing and store reusable signals and targets.--prepare
  3. 3Sanity-check locallyRun a small subset to catch dependency, shape, and storage errors quickly.--debug
  4. 4Run the full baselineTrain the published model unchanged and save its score for comparison.

The preparation cache can be the slowest step. A successful debug run checks the full data and model path before you spend time on full training.

Public-data reference

Compare reference scores.

Before the competition opens, these published baselines are the reference scores to reproduce and beat while validating your pipeline. From September 21, competition submissions and leaderboard scores become the official comparison.

Track Baseline Family Params Train (GPU-h) Reference score Configuration
Track 01 ChanceEEG-to-Image reference 0 0 Top-5 2.22 ± 0.31 Not applicable
Track 01 EEGNetEEG-to-Image · Lawhern et al., 2018 CNN 0.04 M 2 Top-5 28.13 ± 0.14 configs/img/eegnet.yaml
Track 01 REVEEEG-to-Image · Elouahidi et al., 2025 Foundation 14 M 0.5 (probe) Top-5 84.75 ± 0.38 configs/img/reve_frozen.yaml
Track 02 ChanceBCI decoding reference 0 0 Bal. Acc 24.81 ± 1.03 Not applicable
Track 02 EEGNetBCI decoding CNN 0.04 M 4 Bal. Acc 58.58 ± 0.34 configs/bci/eegnet.yaml
Track 02 REVEBCI decoding Foundation 14 M 1 (probe) Bal. Acc 68.04 ± 0.73 configs/bci/reve.yaml
Track 03 ChanceSleep onset reference 0 0 W-bMAE 205.42 ± 0.01 s Not applicable
Track 03 EEGNet-sleepSleep onset CNN 0.04 M 4 W-bMAE 143.30 ± 0.40 s configs/sleep/eegnet.yaml
Track 03 REVE-sleepSleep onset Foundation 14 M 1 (probe) W-bMAE 134.89 ± 2.02 s configs/sleep/reve.yaml
Track 04 NeuroPoseEMG-to-Pose · Liu et al., 2021 Pose CNN - paper MAE 17.5 ± 1.5° tasks/emg/pose/config.yaml
All values mean ± std from NeuralBench (Banville et al., 2026), Table 1. Sleep metric is W-bMAE (weighted-binned MAE) per the proposal §1.4. Register on Codabench →
Competition handoff

Package and submit.

Your preparation workflow becomes a competition submission when the portal opens. The official NeuralBench submission guide is the source for the current package format and upload instructions.

Before Sep 21

Build a verified local pipeline.

  1. Use the public seed datasets and your track guide.
  2. Reproduce a baseline and check it against the reference score.
  3. Train your own model in NeuralBench or with your own compatible pipeline.
From Sep 21

Submit for hidden evaluation.

  1. Use the released competition data and final track instructions.
  2. Package your trained weights and a minimal inference script in the required format.
  3. Upload the package to the track portal. Organizers run it on the hidden test set.
Read the NeuralBench submission guide ↗ Choose a track and register →

What changes at launch: the public start kits remain useful for development, while the competition release, portal instructions, and hidden evaluation create the official ranking. Check the submission guide again before packaging your final model.