metadata
dataset_info:
features:
- name: image
dtype: image
- name: source
dtype: string
splits:
- name: train
num_bytes: 9077485136.748
num_examples: 46684
download_size: 8183528102
dataset_size: 9077485136.748
configs:
- config_name: default
data_files:
- split: train
path: data/train-*
This is a sample cache dataset for caching the image activations for interpreting the visual features in the sae. It is being curated using the following script
We select these data to curate a relatively small size but more diverse cache dataset to interpret the features. For more information, you can refer to the GitHub