NeuroConText checkpoint
The CLIP head of NeuroConText (Ghayem et al., MICCAI 2024; Imaging Neuroscience 2026) as a PyTorch state dict
(best_val.pt, 27 tensors, 14.7 MB): a text tower on Mistral-7B-v0.1 embeddings (projection to 512, two residual
blocks) and an image tower on DiFuMo-512 features (three residual blocks), with logit_scale.
The NeuroConText authors publish their code (github.com/ghayem/NeuroConText) and data (Zenodo record 14169410,
CC-BY-4.0) but no trained weights. This file is the best_val.pt written by their training script (training.py,
the epoch with the lowest validation loss) on that data, trained by the NeuroVLM authors in 2025-2026 and shared here
so that NeuroVLM's Fig. S3 (recall@k of NeuroVLM, NiCLIP and NeuroConText on the PubMed test papers in neither
baseline's training set) can be reproduced from fetched inputs. Its text pipeline (text_preprocessing_pipeline.pkl,
L2 normalisation then a StandardScaler) and training PMIDs come from the Zenodo archive; NeuroVLM's
scripts/data/fetch_baselines.py reads them from it.
sha256 0ceae4608bf8fc93856dd6b5eff24dcd5cbc97c0806f0eca34b149a538e19c30. Pinned in NeuroVLM's
scripts/config/hf_manifest.json and loaded by scripts/figures/main/figS3_recall_comparison.py (the architecture is
copied there from layers.py of the NeuroConText repository). The NeuroConText repository carries no license file;
these weights are shared for reproducibility of the comparison only.