Ir para o conteúdo
ModelListadosExterno
Baixar

Imagenet Sae Top K 64 Patches Only Layer 7 Hook Resid Post 128 79

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-128-79/tree/34c969ce6b396c463516ac15d604ea3de23165a3
TipoOtherVersãoRevision 34c969cCnota

Fonte do modelo

Tipo
Other
Tamanho do arquivo
288 MB
Versão
vRevision 34c969c
Atualizado
6 de out. de 2026
Downloads
7
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: 0.0001
  • L1 Coefficient: 0.0002
  • Batch Size: 4096
  • Context Size: 49

Performance Metrics

Sparsity

  • L0 (Active Features): 128.0000
  • Dead Features: 0
  • Mean Log10 Feature Sparsity: -2.8098
  • Features Below 1e-5: 0
  • Features Below 1e-6: 0
  • Mean Passes Since Fired: 0.0573

Reconstruction

  • Explained Variance: 0.7958
  • Explained Variance Std: 0.0439
  • MSE Loss: 0.0024
  • L1 Loss: 0
  • Overall Loss: 0.0024

Training Details

  • Training Duration: 4202 seconds
  • Final Learning Rate: 0.0000
  • Warm Up Steps: 500
  • Gradient Clipping: 1

Additional Information

Fontes

1 fonte
Hugging FacePrimárioVersão 34c969ce6b396c463516ac15d604ea3de23165a3
huggingface.co/Prisma-Multimodal/imagenet-sae-top_k-64-patches_only-layer_7-hook_resid_post-128-79/tree/34c969ce6b396c463516ac15d604ea3de23165a3

Verificado 6 de out.

Artefatos de modelo

1 artefato

weights.pt

pt · 288 MB · SHA-256 223ec87a2c4a…a6dc · Hugging Face

Baixar

Trechos de fonte

3 trechos
Readme29 de jul.huggingface.co/Prisma-Multimodal/imagenet-sae-top_k-64-patches_only-layer_7-hook_resid_post-128-79

# 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**: 0.0001 - **L1 Coefficient**: 0.0002 - **Batch Size**: 4096 - **Context Size**: 49 ## Performance Metrics ### Sparsity - **L0 (Active Features)**: 128.0000 - **Dead Features**: 0 - **Mean Log10 Feature Sparsity**: -2.8098 - **Features Below 1e-5**: 0 - **Features Below 1e-6**: 0 - **Mean Passes Since Fired**: 0.0573 ### Reconstruction - **Explained Variance**: 0.7958 - **Explained Variance Std**: 0.0439 - **MSE Loss**: 0.0024 - **L1 Loss**: 0 - **Overall Loss**: 0.0024 ## Training Details - **Training Duration**: 4202 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/165d9ce1-tinyclip_sae_16_hyperparam_sweep_lr/n_images_1300020.pt - **Wandb Run**: https://wandb.ai/perceptual-alignment/imagenet-sweep-topk-patches_all_layers/runs/zdtnf9b0 - **Random Seed**: 42

Note6 de out.huggingface.co/Prisma-Multimodal/imagenet-sae-top_k-64-patches_only-layer_7-hook_resid_post-128-79

Source context: 7 downloads · 0 likes · Repo Prisma-Multimodal/imagenet-sae-top_k-64-patches_only-layer_7-hook_resid_post-128-79

Note29 de jul.huggingface.co/Prisma-Multimodal/imagenet-sae-top_k-64-patches_only-layer_7-hook_resid_post-128-79

Source context: 5 downloads · 0 likes · Repo Prisma-Multimodal/imagenet-sae-top_k-64-patches_only-layer_7-hook_resid_post-128-79