EEG-HealthLLM

EEG-HealthLLM

EEG-HealthLLM provides a unified closed-set language interface for five heterogeneous EEG health assessment tasks: epileptic seizure detection, major depressive disorder assessment, anxiety-risk assessment, depressive-symptom risk assessment, and emotional-state assessment.

This repository contains the complete released model extension:

• the shared language-side adapter;
• five task-specific sequence/global projectors;
• task manifests and portable route configuration.

The frozen microsoft/Phi-3-mini-4k-instruct backbone is not redistributed here.

Code

The model uses custom EEG routing and language-adapter code. Use the official implementation at:

https://github.com/maruisaa/EEG-HealthLLM

This release is not intended to be loaded through the standard Hugging Face text-generation pipeline without the accompanying code.

Download

hf download maruisaa/EEG-HealthLLM \
  --local-dir "weights/EEG-HealthLLM"

Alternatively, use Python:

from huggingface_hub import snapshot_download

snapshot_download(
    repo_id="maruisaa/EEG-HealthLLM",
    local_dir="weights/EEG-HealthLLM",
)

Files

EEG-HealthLLM/
|-- adapter_config.json
|-- adapter_model.bin
|-- config.json
|-- unified_eeg_meta.json
|-- unified_routes_5tasks.json
`-- eeg_adapters/
    |-- epilepsy/
    |-- mumtaz/
    |-- gad7/
    |-- phq9/
    `-- seed/

Each task directory contains manifest.json and projectors/model.safetensors, together with route-specific checkpoint files where applicable.

Quick evaluation

After cloning the code repository and downloading this snapshot into the directory shown above:

PHI=/path/to/Phi-3-mini-4k-instruct bash "weights/evaluate_example.sh" seed

Valid task names are epilepsy, mumtaz, gad7, phq9, and seed.

Reported test performance

Data

Raw EEG recordings are not included. Users must obtain the datasets from their official sources and follow the preprocessing and subject-independent split metadata distributed with the code repository.

Intended use and limitations

EEG-HealthLLM is released for research and reproducibility. Its outputs must not be interpreted as clinical diagnoses, and the model is not intended for autonomous medical decision-making. Performance may change across populations, recording devices, electrode layouts, preprocessing pipelines, and distribution shifts.

Citation

Please cite the accompanying EEG-HealthLLM paper when using these weights. Full bibliographic metadata will be added after publication.