brain-lora

fMRI encoding models from Scaling Parameter-Efficient Fine-Tuning for Brain Encoding. Code: github.com/zfying/brain-lora.

Each model is a LoRA adapter on Qwen/Qwen3-Omni-30B-A3B-Thinking plus a readout that predicts the BOLD response in 1000 Schaefer parcels for the four subjects of the Algonauts 2025 challenge from the video, audio and transcript of a stimulus.

folder	readout	r	sub-01	sub-02	sub-03	sub-05
lora_linear	linear	0.326	0.323	0.330	0.350	0.302
lora_cross_attn	cross-attention	0.331	0.326	0.334	0.356	0.308

r is the mean per-parcel Pearson correlation on Friends season 6 (held out).

Files

• adapter_config.json, adapter_model.safetensors: the LoRA adapter (PEFT format)
• readout.safetensors: the readout weights
• config.json: the training configuration and the readout's shape

Usage

python -m brain_lora.cli.hub download zfying/brain-lora lora_linear \
    --data-dir data --run-dir runs/lora_linear
python -m brain_lora.cli.evaluate runs/lora_linear

See the code repository for setup and data preparation.

Citation

@misc{ying2026scaling,
      title={Scaling Parameter-Efficient Fine-Tuning for Brain Encoding}, 
      author={Zhuofan Josh Ying and Guanglong Sun and Nolan Ho and Kieran Liu and Rui Yang and Andrew Chen and Nikolaus Kriegeskorte and Pony Ma and Hossein Adeli},
      year={2026},
      eprint={2610.26802},
      archivePrefix={arXiv},
      primaryClass={cs.LG},
      url={https://arxiv.org/abs/2610.26802}, 
}