rphammonds commited on
Commit
d103770
·
verified ·
1 Parent(s): b7c288b

Revision 2026-10: final autoencoder (git 51bedf4, hf_manifest 49f3e3a9825b, 2026-10-10)

Browse files

Run: scripts/outputs/runs/final (scripts/pipeline/train_pipeline.py), finished 2026-10-10T19:56:25.
Training code: git 51bedf4ed6c6b2a0868f8f5fcb5bdef2e330e322 with 131 modified files (runs/final/provenance.json).
Inputs: scripts/config/hf_manifest.json sha256 49f3e3a9825b9754f921b12e75575dcd75e015f15faf9b0b94748acaa370b78b.
Published by scripts/pipeline/push_hf_release.py at git e31d9e5182bf059a7078cfc6944b10974dce5a89.

autoencoder.safetensors 151d62eebc805ea79007415355c8b26069af7d9c8ee545b7fddb6a5922731f55

Files changed (2) hide show
  1. README.md +20 -0
  2. autoencoder.safetensors +1 -1
README.md CHANGED
@@ -17,3 +17,23 @@ Autoencoder weights and brain-mask metadata for NeuroVLM anatomical atlas vector
17
  - `config.json`: architecture and file metadata for downstream loaders.
18
 
19
  Load through `neurovlm.models.load_model("autoencoder")` after the package loader update.
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
17
  - `config.json`: architecture and file metadata for downstream loaders.
18
 
19
  Load through `neurovlm.models.load_model("autoencoder")` after the package loader update.
20
+
21
+ ## Revision 2026-10
22
+
23
+ Weights of the revised NeuroVLM paper. What changed:
24
+
25
+ - **Deterministic training.** Every stage runs with seeded data loaders and deterministic kernels. Two runs from scratch gave byte-identical weights on the same GPU model, driver and library versions.
26
+ - **MLP projector.** The brain-to-text projectors (stages 13 and 14) are residual MLPs (28.3 M parameters) instead of Q-Formers. A package names its projector in `config.json["projector"]`.
27
+ - **Stage-14 recipe.** Each training text is projected onto the canonical bank of its own level, every training map is re-binarized per batch, and title tokens weigh 5 in the loss. Five seeds were trained and the one with the lowest validation loss (seed 3) is packaged.
28
+ - **Calibration.** Text-to-brain maps are multiplied by one global factor per model, fitted on the validation split and clipped at 1 (`map_calibration.json`, `head` 19.16 and `adapter` 6.13).
29
+ - **Title decoding.** After a [NETWORK] or [REGION] tag the decoder writes one of the canonical names in `title_vocabulary.json` (48 network and 83 region titles), so it never invents a title.
30
+
31
+ `autoencoder.safetensors` is the retrained autoencoder (stage 08). `mask.npz` and `config.json` are unchanged.
32
+
33
+ | file | size | sha256 |
34
+ |---|---|---|
35
+ | `autoencoder.safetensors` | 239.7 MB | `151d62eebc805ea79007415355c8b26069af7d9c8ee545b7fddb6a5922731f55` |
36
+
37
+ Provenance: `scripts/outputs/runs/final` of NeuroVLM (`scripts/pipeline/train_pipeline.py`, finished 2026-10-10T19:56:25), training code at git `51bedf4` with 131 modified files, inputs pinned by `scripts/config/hf_manifest.json` (sha256 `49f3e3a9825b`). `scripts/README.md` records the later runs from scratch that gave the same bytes.
38
+
39
+ Earlier weights: revision `b7c288b34d68a71dd303a1a9de02c8aadf6dd1ae`.
autoencoder.safetensors CHANGED
@@ -1,3 +1,3 @@
1
  version https://git-lfs.github.com/spec/v1
2
- oid sha256:26b23c79aebdb73910af32b7eaf2d6bd82c8a66bb5e3d46ed907de33b0019acf
3
  size 239712256
 
1
  version https://git-lfs.github.com/spec/v1
2
+ oid sha256:151d62eebc805ea79007415355c8b26069af7d9c8ee545b7fddb6a5922731f55
3
  size 239712256