OHBA Software Library: Foundation Model Toolbox

This repository contains a TensorFlow-based tokeniser and foundation model (MEG-GPT) for parcellated MEG data.

Preprint: https://arxiv.org/abs/2510.18080.

Installation

We recommend using mamba to install osl-foundation, which can be installed with:

wget "https://github.com/conda-forge/miniforge/releases/latest/download/Miniforge3-$(uname)-$(uname -m).sh"
bash Miniforge3-$(uname)-$(uname -m).sh
rm Miniforge3-$(uname)-$(uname -m).sh

Then osl-foundation can be installed with:

git clone https://github.com/OHBA-analysis/osl-foundation.git
cd osl-foundation
mamba env create -f envs/oslf.yml
conda activate oslf
pip install -e .

Note, MEG-GPT requires TensorFlow 2.11 and comes with osl-dynamics (v2.1.8).

BMRC Cluster (Oxford)

conda/mamba are available as a software module:

module load Miniforge3

osl-foundation can be installed with:

git clone https://github.com/OHBA-analysis/osl-foundation.git
cd osl-foundation
mamba env create -f envs/bmrc.yml
conda activate oslf
pip install -e .

Note, the following CUDA module needs to be loaded on BMRC to use TensorFlow:

module load cuDNN/8.4.1.50-CUDA-11.7.0

Usage

See the examples directory.

Load the pre-trained models (tokenizer, MEG-GPT)

First download the model weights (which are hosted on Hugging Face):

git clone https://huggingface.co/OHBA-analysis/MEG-GPT models
cd models
git lfs install --local
git lfs pull

Then the models can be loaded with:

from osl_foundation import load_model

tokenizer = load_model("tokenizer")
meg_gpt = load_model("meg-gpt", checkpoint="latest")