Datasets:
Commit History
8 days ago
-
Fix image order in corpus (mteb#5499)
Verified signatureThe corpus images were assigned to the wrong `img_k` IDs during conversion, so the qrels (`text_k -> img_k`) point to the wrong image. See https://github.com/embeddings-benchmark/mteb/issues/5499. This PR replaces only `corpus-00000-of-00001.parquet`. The images are rebuilt from the authors' source images (`BennoKrojer/ImageCoDe@b7b7a73`, same archive as Zenodo 6518944): within each set, the k-th image in numeric filename order becomes `<set>_img_<k>`, as in the official evaluator. Row order, IDs, schema and the existing image format (256×256 center crop, JPEG q75) are kept. Queries and qrels are unchanged. `mteb/imagecode` and `mteb/imagecode-multi` share the same corpus file, so both get the same replacement. Checked locally with `openai/clip-vit-base-patch16@57c21647` through `mteb.evaluate`: - ImageCoDe `image_acc`: 12.0% → 22.8% (paper, validation: 21.8%) - ImageCoDeT2IRetrieval `hit_rate_at_3`: 9.8% → 16.5% Scripts, pinned inputs and results: https://github.com/lxer19/mteb/b…a35fbeb
-
Duplicate from JamieSJS/imagecode-multi
Co-authored-by: Jamie Stirling <JamieSJS@users.noreply.huggingface.co>0a27b28
