Display-only change: restore the existing banner and current quality-cost figure, use supported inline math delimiters, fix pi/@ tokenization, and update the companion GitHub link. Only README.md and its manifest checksum change; all data splits, features, media, predictions, and experimental results remain unchanged.
Restore dataset-card figures and fix Hugging Face math rendering
Display-only change: restore the existing banner and current quality-cost figure, use supported inline math delimiters, fix pi/@ tokenization, and update the companion GitHub link. Only README.md and its manifest checksum change; all data splits, features, media, predictions, and experimental results remain unchanged.
Yesterday
Standardize multimodal media as train, standard test, and robust test
Rename the three media archives without changing their bytes; update only media_json.archive in multimodal Parquet rows. Add a concise media README and consistently named checksum manifest. Requests, image members, scores, embeddings, prediction arrays, and report metrics are unchanged.
Remove development-only records from the public benchmark package
20,167 exact-match successful score restorations; known failures stay zero. Unverified legacy scores unchanged. Shared CLIP text/all-image late fusion, 12 canonical public baselines and ID-aligned prediction replay for all three scenarios.
