keypa/vision-adapter-images

Vision Adapter Image Corpus

Processed image corpus used to train the Vision-Adapter: 138,987 images — 79,659 agentic UI images ("screenshots" + "multistep" subsets derived from wave-ui-25k, ShowUI-desktop, and aguvis-stage2) plus the embeddable subsets of HuggingFaceM4/the_cauldron used for general/reasoning/conversational fine-tuning.

Each image has been resized to ≤300k pixels and padded to 28-pixel multiples to match MoonViT-V2 preprocessing.

Fields

• image: the encoded image bytes (binary column — JPEG for the agentic subset, PNG for most cauldron rows; the extension is whatever the source shipped, not a guarantee of format).
• filename: source/basename, e.g. agentic/aitw_000000.jpg, cauldron/chartqa-0000001-0.png.
• source: agentic or cauldron.
• size: bytes of the encoded image.

filename is load-bearing

The embedding key is sha1(filename)[:20].pt, so this column is what joins the corpus to the precomputed embeddings in keypa/vision-adapter-manifests. Renaming or reordering files breaks that join.

It is also what the MoonViT patch grid is derived from: the grid is a deterministic function of the image dimensions under the preprocessing contract above, which is how train_manifest_grids.jsonl in the manifests repo carries measured geometry instead of a token-count approximation.

How to use

from datasets import load_dataset

ds = load_dataset("keypa/vision-adapter-images")
print(ds)