Datasets:
Release Notes
release_v1
This release package is cleaned for public publication.
Release-facing changes
- removed the auxiliary canonical split from the public package
- removed paper-era legacy manifests from the public package
- removed machine-local path assumptions from scripts and docs
- kept only official
trainandtestsplit manifests - preserved raw query relevance provenance in
metadata/source/ - preserved gallery-filtered query relevance in
metadata/query_relevance.json - added a minimal
timmevaluation example - kept only the validation utility needed by end users
Official counts
- train gallery:
47,882images /23,941crater IDs - test gallery:
50,000images /25,000crater IDs - test query:
5,000images /1,000crater IDs
Hosting
- primary: Hugging Face Datasets (
jfang/CraterBench-R) - all images distributed as a single
images.ziparchive (1.8 GB) - extract with
unzip images.zipto produceimages/gallery/andimages/query/