Datasets:

---

brennercruvinel

/

mtg-urna-benchmark

Like 2

Merge after discussion #3 (the candidate sidecars), which this card describes.

• queries and qrels configs: 39,927 text-to-image queries of experiments 13, 14, 15 and 20, each "artwork of the card {label}" with exactly one relevant row, the card itself (the benchmark's own evaluation, hit@k, no graded relevance). Written by benchmark/tools/export_queries.py from the tracked lists in benchmark/corpora (GitHub PR #35).
• How to open: hf download now passes --repo-type dataset (without it the command looks for a model repo), plus the SHA256SUMS check. "any urna after 0.4.0" (a version with no tag) is now "urna 0.5.0 and later".
• The two searches kept apart: text on the potion space (retrieve, ask) and text-to-image through a model's text tower; the siglip2 flow as it is now: the snapshot pinned to timm/ViT-B-16-SigLIP2@eee10eff, the urna loader of hoffresearch/urna #271 and #273, offline, no refs/main; remote code only for jina and wemm.
• Files: candidate directories with their new sidecars; crf50 is the retrieval release's file at both paths.
• Not included: the siglip2 weights' source and revision.

GitHub side: PR #36 (README and docs/environment.md).