Text2fMRI

Text2fMRI offers a suite of encoding models, available through the Hugging Face collection 'ShreyDixit/Text2fMRI', designed to predict whole-brain fMRI responses solely from video transcripts.

Multiple configurations are available to suit different resource constraints. The smallest and most lightweight configuration consists of approximately 52M trainable parameters, leveraging a frozen 500M parameter LLM (e.g., Qwen-0.5B) or feature extraction.

Trained on the CNeuroMods dataset (Friends and Movie10)—the same data used for the Algonauts 2025 Challenge—this model generates in silico neural responses without requiring visual or audio inputs. Despite its efficiency, even the smallest model outperforms standard baselines and achieves near-SOTA performance in auditory and language-selective cortices.