docs(weights): document trained model provenance and checksums

5be1845 verified 18 days ago

2.03 kB

---
tags:
- fmri
- nsd
- reproduction
- subject-1
---

# Brain-Diffuser regression — Subject 1 reproduction weights

Three self-fitted Subject1 regression heads from the original Brain-Diffuser recipe: VDVAE(alpha50000), CLIP vision(alpha60000) and CLIP text(alpha100000). The pretrained VDVAE/Versatile Diffusion/CLIP backbones are downloaded dependencies and are not republished.

The input is the native nsdgeneral ROI vector with15724voxels in the original C-order. Divide brain values by300, then apply stored training mean and std(ddof1). Each pickle contains `weight` and `bias`. CLIP weights retain the original author's float32 serialization; VDVAE retains its original serialization. CLIP target-normalization files contain only aggregate training-feature mean/std(ddof0), computed from the already-frozen training features, with no refitting.

After regression, the original pipeline z-scores predictions over the entire982test pool before applying training-target mean/std. This is a batch-dependent, transductive calibration: applying it independently to a smaller subset changes results. VDVAE normalization also preserves the original full982prediction statistics. Reconstructing from the serialized float32 CLIP coefficients can differ slightly from original in-memory regression predictions; no new bit-exact reconstruction claim is made. Frozen models, NSD inputs and per-sample target features are not part of this release.

Upstream: https://github.com/ozcelikfu/brain-diffuser .

## Provenance and integrity

Source revision: `5a8b2e62a32a00b136c620fdb3f30ea797744c8e`. These repositories preserve our trained artifacts. `manifest.json` and `SHA256SUMS` list exact file sizes/hashes. This is an inference-weight release; full optimizer/RNG continuation states remain in the local project. The method name credits the original authors; this account is not the official upstream release. Use of upstream architectures and pretrained dependencies remains subject to their original licenses.

	---
	tags:
	- fmri
	- nsd
	- reproduction
	- subject-1
	---

	# Brain-Diffuser regression — Subject 1 reproduction weights

	Three self-fitted Subject1 regression heads from the original Brain-Diffuser recipe: VDVAE(alpha50000), CLIP vision(alpha60000) and CLIP text(alpha100000). The pretrained VDVAE/Versatile Diffusion/CLIP backbones are downloaded dependencies and are not republished.

	The input is the native nsdgeneral ROI vector with15724voxels in the original C-order. Divide brain values by300, then apply stored training mean and std(ddof1). Each pickle contains `weight` and `bias`. CLIP weights retain the original author's float32 serialization; VDVAE retains its original serialization. CLIP target-normalization files contain only aggregate training-feature mean/std(ddof0), computed from the already-frozen training features, with no refitting.

	After regression, the original pipeline z-scores predictions over the entire982test pool before applying training-target mean/std. This is a batch-dependent, transductive calibration: applying it independently to a smaller subset changes results. VDVAE normalization also preserves the original full982prediction statistics. Reconstructing from the serialized float32 CLIP coefficients can differ slightly from original in-memory regression predictions; no new bit-exact reconstruction claim is made. Frozen models, NSD inputs and per-sample target features are not part of this release.

	Upstream: https://github.com/ozcelikfu/brain-diffuser .

	## Provenance and integrity

	Source revision: `5a8b2e62a32a00b136c620fdb3f30ea797744c8e`. These repositories preserve our trained artifacts. `manifest.json` and `SHA256SUMS` list exact file sizes/hashes. This is an inference-weight release; full optimizer/RNG continuation states remain in the local project. The method name credits the original authors; this account is not the official upstream release. Use of upstream architectures and pretrained dependencies remains subject to their original licenses.