Download evaluation/human_metrics.json from AhmedSSabir/BERT-CNN-Reward-Caption-Reranker: direct link, hf CLI and curl.
- Browser
- Download file 1.76 kB
-
https://huggingface.co/AhmedSSabir/BERT-CNN-Reward-Caption-Reranker/resolve/main/evaluation/human_metrics.json
- Command line
-
hf download hf://AhmedSSabir/BERT-CNN-Reward-Caption-Reranker/evaluation/human_metrics.json
-
curl -L -o human_metrics.json https://huggingface.co/AhmedSSabir/BERT-CNN-Reward-Caption-Reranker/resolve/main/evaluation/human_metrics.json
1.76 kB
| { | |
| "dataset": "Original Flickr8K ExpertAnnotations.txt", | |
| "pairs": 5822, | |
| "images": 1000, | |
| "checkpoint_sha256": "125a083e3258b7ee038b6fb3e9d40a2d6078e2a325fbc1c4ba9cd062e674daaf", | |
| "metrics": { | |
| "clipscore_vit_b32": { | |
| "examples": 5822, | |
| "spearman": 0.7273458253579359, | |
| "kendall_tau_b": 0.5760866201430647, | |
| "kendall_tau_c": 0.5631096474308253, | |
| "pearson": 0.7484619266030074 | |
| }, | |
| "bert_cnn_resnet152": { | |
| "examples": 5822, | |
| "spearman": 0.23897661511730053, | |
| "kendall_tau_b": 0.17511563484276757, | |
| "kendall_tau_c": 0.17116331913460656, | |
| "pearson": 0.20429245894996126, | |
| "normalized_human_mae": 0.47293288321293026, | |
| "empty_image_contexts": 218 | |
| }, | |
| "bert_cnn_clip_imagenet": { | |
| "examples": 5822, | |
| "spearman": 0.23477193485123116, | |
| "kendall_tau_b": 0.1718365123138481, | |
| "kendall_tau_c": 0.1679471771227406, | |
| "pearson": 0.19951677072314783, | |
| "normalized_human_mae": 0.5044658286184057, | |
| "empty_image_contexts": 386 | |
| }, | |
| "bert_cnn_combined": { | |
| "examples": 5822, | |
| "spearman": 0.3646498307040494, | |
| "kendall_tau_b": 0.2680415373382081, | |
| "kendall_tau_c": 0.2620033341773608, | |
| "pearson": 0.3111999615665536, | |
| "normalized_human_mae": 0.4910941707480373, | |
| "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" | |
| ] | |
| } | |