NeuroMM 2026 Track 3: EEGBind checkpoints
These are the six selected inference checkpoints for the NeuroMM 2026 NMM-Source-IED code release. The five fold checkpoints are combined by averaging log probabilities. eeg_anchor.pt is the frozen EEG backbone used to extract candidate EEG features.
The repository contains only Track 3 weights. It contains no challenge data, labels, submission predictions, or checkpoints from other tracks.
| File | Purpose |
|---|---|
eeg_anchor.pt |
Frozen STEEGFormer EEG feature extractor |
folds/fold_0.pt through folds/fold_4.pt |
Five EEG-video classifier folds |
MANIFEST.tsv |
SHA-256 checksums for all six files |
Use the code release to verify and run these checkpoints:
hf download neuro1104/NeuroMM2026-IED-Detection --local-dir /path/to/track3-weights
python scripts/prepare_weights.py --source /path/to/track3-weights --verify
bash scripts/run_inference.sh
The selected challenge submission achieved 0.8395 public leaderboard weighted F1. Full inference requires the official NeuroMM candidate data. The earlier training initialization checkpoint and full training chain are not part of this inference release.
The complete codebase underlying this work will be made publicly available in November 2026.