SAMBA-100s

Project website · Code · Paper · SAMBA-2s

SAMBA-100s was pretrained on 100-second windows from the Temple University Hospital Abnormal EEG Corpus (TUAB), using 12,800 samples at 128 Hz and 16 bipolar channels. Retraining followed the original method with a rebuilt TUAB pipeline and validation split. The encoder supports other window lengths and electrode layouts.

Download

Download config · Download weights

Keep both files in the same directory. To download them together from a terminal:

pip install -U huggingface_hub
hf download Dcode-AI/SAMBA-100s config.json model.safetensors --local-dir ./SAMBA-100s

Usage

Install SAMBA in a CUDA environment:

git clone https://github.com/Jiazhen-Hong/SAMBA.git
cd SAMBA
pip install -e '.[mamba,hub]'

Load the downloaded directory. If you skipped the download step, the following example fetches both files automatically:

import torch
from huggingface_hub import snapshot_download
from samba_eeg import SambaModel

folder = snapshot_download(
    "Dcode-AI/SAMBA-100s",
    allow_patterns=["config.json", "model.safetensors"],
)
model = SambaModel.from_pretrained(folder).cuda().eval()

For files you already downloaded, pass their local directory to SambaModel.from_pretrained in place of folder.

Prepare EEG at 128 Hz. Use signals shaped [batch, channels, samples] and electrode positions shaped [channels, 3] or [batch, channels, 3]. Match the signal channel order and use the same coordinate frame and units as the target positions in config.json. TUAB pretraining used a 0.5–45 Hz bandpass. The decoder reconstructs EEG on a fixed 16-channel bipolar montage.

signals = signals.to(device="cuda", dtype=torch.float32)
positions = positions.to(device="cuda", dtype=torch.float32)

# Skip these two lines if the signals are already normalized this way.
scale = torch.quantile(signals.abs(), 0.95, dim=-1, keepdim=True)
signals = signals / scale.clamp_min(1e-8)

with torch.inference_mode():
    features = model.encode(signals, positions)
    pooled_features = features.mean(dim=1)

features has shape [batch, floor(samples / 8), 128]. Use pooled_features for linear probing.

Learned EEG spatial weights

SAMBA uses electrode coordinates to learn how to combine EEG channels. These maps show how the spatial weights change during training.

Eyes open vs. closed Driver distraction Motor imagery
Eyes open vs. closed Driver distraction Motor imagery

Citation

@inproceedings{hong2026samba,
  title     = {{SAMBA}: Toward an {EEG} Foundation Model via Differential Mamba and Coordinate-Driven Spatial Embedding},
  author    = {Hong, Jiazhen and Mackellar, Geoffrey and Ghane, Soheila},
  booktitle = {NeurReps 2026: Symmetry and Geometry in Neural Representations (Proceedings Track)},
  year      = {2026}
}
Downloads last month
19
Safetensors
Model size
1.06M params
Tensor type
F32
·
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support

Collection including Dcode-AI/SAMBA-100s

Paper for Dcode-AI/SAMBA-100s