Ir para o conteúdo
ModelListadosExterno
Baixar

Imagenet Sae Top K 64 Patches Only Layer 7 Hook Resid Post 64 78

This model is a sparse autoencoder trained on CLIP's internal representations.

Fontehuggingface.co/Prisma-Multimodal/imagenet-sae-top_k-64-patches_only-layer_7-hook_resid_post-64-78/tree/3b4a260c89d80187614f128de0d9d4db847b99ce
TipoOtherVersãoRevision 3b4a260Cnota

Fonte do modelo

Tipo
Other
Tamanho do arquivo
288 MB
Versão
vRevision 3b4a260
Atualizado
7 de out. de 2026
Downloads
3
Fonte
Hugging Face

Descrição da fonte

CLIP Sparse Autoencoder Checkpoint

This model is a sparse autoencoder trained on CLIP's internal representations.

Model Details

Architecture

  • Layer: 7
  • Layer Type: hook_resid_post
  • Model: open-clip:laion/CLIP-ViT-B-32-DataComp.XL-s13B-b90K
  • Dictionary Size: 49152
  • Input Dimension: 768
  • Expansion Factor: 64
  • CLS Token Only: False

Training

  • Training Images: 1299936
  • Learning Rate: 0.0005
  • L1 Coefficient: 0.0002
  • Batch Size: 4096
  • Context Size: 49

Performance Metrics

Sparsity

  • L0 (Active Features): 64.0000
  • Dead Features: 0
  • Mean Log10 Feature Sparsity: -3.2880
  • Features Below 1e-5: 6
  • Features Below 1e-6: 0
  • Mean Passes Since Fired: 0.4979

Reconstruction

  • Explained Variance: 0.7831
  • Explained Variance Std: 0.0517
  • MSE Loss: 0.0025
  • L1 Loss: 0
  • Overall Loss: 0.0025

Training Details

  • Training Duration: 4076 seconds
  • Final Learning Rate: 0.0000
  • Warm Up Steps: 500
  • Gradient Clipping: 1

Additional Information

Fontes

1 fonte
Hugging FacePrimárioVersão 3b4a260c89d80187614f128de0d9d4db847b99ce
huggingface.co/Prisma-Multimodal/imagenet-sae-top_k-64-patches_only-layer_7-hook_resid_post-64-78/tree/3b4a260c89d80187614f128de0d9d4db847b99ce

Verificado 7 de out.

Artefatos de modelo

1 artefato

weights.pt

pt · 288 MB · SHA-256 f965dc06f18a…5944 · Hugging Face

Baixar

Trechos de fonte

3 trechos
Readme30 de jul.huggingface.co/Prisma-Multimodal/imagenet-sae-top_k-64-patches_only-layer_7-hook_resid_post-64-78

# CLIP Sparse Autoencoder Checkpoint This model is a sparse autoencoder trained on CLIP's internal representations. ## Model Details ### Architecture - **Layer**: 7 - **Layer Type**: hook_resid_post - **Model**: open-clip:laion/CLIP-ViT-B-32-DataComp.XL-s13B-b90K - **Dictionary Size**: 49152 - **Input Dimension**: 768 - **Expansion Factor**: 64 - **CLS Token Only**: False ### Training - **Training Images**: 1299936 - **Learning Rate**: 0.0005 - **L1 Coefficient**: 0.0002 - **Batch Size**: 4096 - **Context Size**: 49 ## Performance Metrics ### Sparsity - **L0 (Active Features)**: 64.0000 - **Dead Features**: 0 - **Mean Log10 Feature Sparsity**: -3.2880 - **Features Below 1e-5**: 6 - **Features Below 1e-6**: 0 - **Mean Passes Since Fired**: 0.4979 ### Reconstruction - **Explained Variance**: 0.7831 - **Explained Variance Std**: 0.0517 - **MSE Loss**: 0.0025 - **L1 Loss**: 0 - **Overall Loss**: 0.0025 ## Training Details - **Training Duration**: 4076 seconds - **Final Learning Rate**: 0.0000 - **Warm Up Steps**: 500 - **Gradient Clipping**: 1 ## Additional Information - **Original Checkpoint Path**: /network/scratch/p/praneet.suresh/celeba_checkpoints/19db19f1-tinyclip_sae_16_hyperparam_sweep_lr/n_images_1300020.pt - **Wandb Run**: https://wandb.ai/perceptual-alignment/imagenet-sweep-topk-patches_all_layers/runs/lmgjyhkd - **Random Seed**: 42

Note7 de out.huggingface.co/Prisma-Multimodal/imagenet-sae-top_k-64-patches_only-layer_7-hook_resid_post-64-78

Source context: 3 downloads · 0 likes · Repo Prisma-Multimodal/imagenet-sae-top_k-64-patches_only-layer_7-hook_resid_post-64-78

Note30 de jul.huggingface.co/Prisma-Multimodal/imagenet-sae-top_k-64-patches_only-layer_7-hook_resid_post-64-78

Source context: 4 downloads · 0 likes · Repo Prisma-Multimodal/imagenet-sae-top_k-64-patches_only-layer_7-hook_resid_post-64-78