angular-separation-vit — reproduction checkpoints
Final (_last.pt) checkpoints backing "Dos modelos idénticos, dos
podas distintas: la dirección dominante de $W_O$ bajo intervención
libre, suave y dura" ("Same function, different pruning: the
dominant direction of $W_O$ under free, soft, and hard intervention").
These are verification artifacts for reproducing the paper's tables —
not meant as general-purpose pretrained weights.
Code: https://github.com/mmunozpl/angular-separation-vit
CSV results derived from these checkpoints: https://huggingface.co/datasets/ManPla/angular-separation-vit-results
Paper (both languages): paper/angular_separation_paper.pdf
(English, canonical for citation) /
paper/separacion_angular_paper.pdf
(Spanish original).
Layout
vitb_clean/attnA_{base,blanda,dura}_seed{42..46}_last.pt # ancla, n=5
vitl_clean/attnA_{base,blanda,dura}_seed{42..44}_last.pt # n=3
dinov2_clean/attnA_base_seed42_last.pt # frozen, n=1
base = fine-tuned ViT with no angular regularizer. blanda = soft
angular probe (R_div, eq. 5). dura = hard variant, reprojected to
the spherical code after every optimizer step. DINOv2 is analyzed
frozen (no fine-tuning), hence a single realization.
All backbones derive from timm pretrained weights, fine-tuned on
ImageNet-100 for the two supervised columns. Upstream pretraining
license, checked against timm's own PretrainedCfg metadata (not
assumed):
| Backbone | Upstream checkpoint | License |
|---|---|---|
| ViT-B/16 | timm/vit_base_patch16_224.augreg2_in21k_ft_in1k |
Apache-2.0 |
| ViT-L/16 | timm/vit_large_patch16_224.augreg_in21k_ft_in1k (Google AugReg) |
Apache-2.0 |
| DINOv2-L | timm/vit_large_patch14_dinov2.lvd142m (Meta) |
Apache-2.0 |
All three are Apache-2.0 at the source, consistent with this repository's license — no upstream restriction carries over.
License
Apache-2.0.