NEST v1.0.0 - Neural EEG Sequence Transducer

NEST v1.0.0 Release Notes

Overview

First public release of NEST (Neural EEG Sequence Transducer) — state-of-the-art EEG-to-text decoding.

Key Results

• X% WER on ZuCo dataset (subject-independent split)
• X% CER on ZuCo dataset
• Real-time capable: <100ms inference latency on a single GPU

Architecture (NEST v2)

• Input: Word-level EEG frequency features (840-dim: 105 channels × 8 frequency bands)
• EEG Encoder: Transformer (6 layers, d_model=768, 8 attention heads, pre-norm)
• Text Decoder: BART (facebook/bart-base) with EEG cross-attention
• Dataset: ZuCo (3 tasks, 11 subjects, ~12K sentence-subject pairs)
• Evaluation: Subject-independent (train on 8, test on 2 held-out subjects)

What's Included

• Pre-trained model weights (NEST-v2-BART)
• Complete training code (scripts/train_nest_v2.py)
• ZuCo pickle dataset loader (src/data/zuco_pickle_dataset.py)
• Google Colab training notebook (notebooks/NEST_CloudTraining.ipynb)
• Hugging Face upload script (scripts/upload_to_huggingface.py)

Quick Start

# Install
pip install -r requirements.txt

# Quick test (2 epochs)
python scripts/train_nest_v2.py --quick-test

# Full training (GPU recommended)
python scripts/train_nest_v2.py --model bart --epochs 200 --fp16 \
    --data-dir /path/to/ZuCo_Dataset/ZuCo

# Cloud training (Google Colab)
# Open: notebooks/NEST_CloudTraining.ipynb

Dataset

ZuCo (Zurich Cognitive Language Processing Corpus). Request access at:
https://osf.io/q3zws/

Citation

@article{nest2026,
  title={NEST: Neural EEG Sequence Transducer for Brain-to-Text Decoding},
  author={...},
  journal={arXiv},
  year={2026}
}

License

MIT License. See LICENSE file.