{ "dataset": "Original Flickr8K ExpertAnnotations.txt", "pairs": 5822, "images": 1000, "checkpoint_sha256": "13502ce4a3578a94effd914525935a2978facd1c7cb243526c3161cbd12db5c2", "metrics": { "examples": 5822, "spearman": 0.2936773447142244, "kendall_tau_b": 0.21296227698549494, "kendall_tau_c": 0.2081649490635248, "pearson": 0.37538085539793303, "normalized_human_mae": 0.22123509002405464, "empty_image_contexts": 0 }, "protocol": "Original mean human ratings; three fixed image-only context policies; no tuning on human ratings", "limitations": [ "Visual extractors differ from training-data extraction; not original-paper replication", "Normalized-human MAE is descriptive; binary-trained outputs are not calibrated human ratings", "Image-context ablations share test data and are not independent validation sets" ] }