Datasets:

---

Duplicated from  JamieSJS/imagecode

mteb

/

imagecode

Like 0

Follow

Massive Text Embedding Benchmark 1.23k

The 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/blob/0d112a0f48ff29364400411bcecc10ba96554f28/scripts/validation/imagecode_5499/README.md