arushsinghal/setu-neural-eeg

SETU Neural EEG

Trained EEG decoders, adapted neural backbones and an auditable imagined-speech evaluation pipeline.

Models on Hugging Face · Research report · All results · Licenses and provenance

This release contains 919 inference checkpoints (5.86 GB) from the frozen experiment inventory, with source, configurations, training histories, evaluation reports and SHA-256 hashes. Checkpoint counts include participant-specific fits, probes, controls and neural models; they are not counts of distinct architectures or foundation models.

SETU trained compact task models and classical estimators and fine-tuned pretrained EEG backbones. We did not pretrain the upstream foundation architectures. The strict validation-selected 32-channel SVM achieved 30.4% held-out accuracy across 15 calibrated participants / 750 trials, with a participant-bootstrap 95% interval of 27.47–33.47%, versus 20% uniform chance. Separate three-second reference and Chisco results are clearly labeled in the research report.

Quick start: released primary SVM

git clone https://github.com/arushsinghal/setu-models.git
cd setu-models
python3 -m venv .venv
source .venv/bin/activate
pip install -r requirements.txt
python release/download.py --primary
python release/predict.py --checkpoint checkpoints/strict_svm_distributed_32_sub-01.joblib --input trial.npy

trial.npy must be measured EEG in microvolts, shape (64, 795), sampled at 256 Hz, in the original dataset electrode order and epoch timing (starts -0.5 seconds relative to imagination onset). The checkpoint stores channel names and selected indices. Labels are Hello, Help me, Stop, Thank you, Yes. This participant-01 model is calibrated for that benchmark participant; it is not a universal new-user decoder. Predictions do not take ground-truth labels as inputs. Probabilities are model scores, not clinically validated confidence.

python release/download.py --all downloads every eligible release checkpoint and pinned runtime asset to the machine where you run it. For this publication, artifacts were transferred directly from remote storage to the hosting services; models/data were not downloaded to the publisher's laptop. Neural loaders use the archived model architecture and runtime assets; reference3s_* uses its separate preprocessing. See reproduction instructions.

Contents

• scripts/: SETU training, preprocessing, adaptation, inference and evaluation code.
• external/CBraMod/models/: pinned upstream MIT runtime source with its copyright notice.
• experiments/: archived training/validation records; reports/: final evaluation and audit records.
• release/: download, inference and verification entry points.
• release-manifest.json: exact checkpoint paths, byte counts, hashes and artifact licenses.

The GitHub repository contains source and reports. Weights and pinned upstream runtime assets live on Hugging Face. CSBrain weights/probes and upstream code are withheld pending redistribution permission; Chisco data/trained artifacts are withheld pending dataset-license clarification. Results for those experiments remain visible. Raw EEG, cached signals, optimizer snapshots and cloud maintenance files are excluded.

Licenses

SETU code is MIT. CBraMod-derived weights retain Apache-2.0 terms; REVE derivatives retain the REVE Responsible Use License v1.0, including use restrictions. The entire weight collection is not represented as unrestricted open source. Consult NOTICE, per-artifact mappings and the bundled licenses before use.

Research scope

This is a cued, retrospective, five-command classification study. It does not establish open-vocabulary speech decoding, neural specificity, new-person generalization, prospective wearable performance or clinical utility. Controls, unsuccessful experiments, the initial seven-label preview exception and pretraining-overlap uncertainty are disclosed in RESEARCH.