metadata
pretty_name: GPIC Final CLIP Embedding Reference Banks
license: mit
size_categories:
- 100K<n<1M
tags:
- image
- embeddings
- clip
- vision-transformer
- mechanistic-interpretability
- retrieval
source_datasets:
- extended
CLIP-GPIC-embeddings
Vision Transformer embeddings for split:TEST (1M) of stanford-vision-lab/gpic
Mainly for use with cross-attention read/no-read bridge experiments from my github.
- Features: pre-indexed image TAR member byte offset and size -> HTTP Range requests to fetch only selected image bytes.
- Meaning: Won't require downloading 1M images for retrieval, will just fetch matches from remote shard.
The MIT license applies to the embedding-bank files, metadata, and supporting code released in this repository to the extent the authors hold rights in those materials. Source GPIC images are not redistributed and remain subject to the GPIC dataset terms. Model implementations and weights remain subject to their respective licenses.