metadata
language: en
tags:
- clip
- vision
- transformers
- interpretability
- sparse autoencoder
- sae
- mechanistic interpretability
license: apache-2.0
library_name: torch
pipeline_tag: feature-extraction
metrics:
- type: explained_variance
value: 81.7
pretty_name: Explained Variance %
range:
min: 0
max: 100
- type: l0
value: 249.88
pretty_name: L0
CLIP-B-32 Sparse Autoencoder x64 vanilla - L1:0.0001
Training Details
- Base Model: CLIP-ViT-B-32 (LAION DataComp.XL-s13B-b90K)
- Layer: 1
- Component: hook_mlp_out
Model Architecture
- Input Dimension: 768
- SAE Dimension: 49,152
- Expansion Factor: x64 (vanilla architecture)
- Activation Function: ReLU
- Initialization: encoder_transpose_decoder
- Context Size: 50 tokens
Performance Metrics
- L1 Coefficient: 0.0001
- L0 Sparsity: 249.8799
- Explained Variance: 0.8170 (81.70%)
Training Configuration
- Learning Rate: 0.0004
- LR Scheduler: Cosine Annealing with Warmup (200 steps)
- Epochs: 10
- Gradient Clipping: 1.0
- Device: NVIDIA Quadro RTX 8000
Experiment Tracking:
- Weights & Biases Run ID: ob776mv6
- Full experiment details: https://wandb.ai/perceptual-alignment/clip/runs/ob776mv6/overview
- Git Commit: e22dd02726b74a054a779a4805b96059d83244aa
Citation
@misc{2024josephsparseautoencoders,
title={Sparse Autoencoders for CLIP-ViT-B-32},
author={Joseph, Sonia},
year={2024},
publisher={Prisma-Multimodal},
url={https://huggingface.co/Prisma-Multimodal},
note={Layer 1, hook_mlp_out, Run ID: ob776mv6}
}