dataset,dataset_name,protocol,notes,metric,metric_name,default_metric,is_default_metric
awa2,AwA2,MLP multi-label attribute probe,"85 animal attributes are predicted from frozen embeddings, the reported score is macro F1.",macro_f1,Macro F1,macro_f1,True
bioscan5m,BIOSCAN-5M,AMI clustering,"Embeddings are reduced with UMAP-50, L2-normalized, clustered and scored with adjusted mutual information.",ami_score,AMI score,ami_score,True
butterflies,Butterflies and Moths Austria,k-NN classification,Species-level k-NN evaluation over frozen embeddings.,val_macro_f1,Validation macro F1,test_macro_f1,False
butterflies,Butterflies and Moths Austria,k-NN classification,Species-level k-NN evaluation over frozen embeddings.,test_macro_f1,Test macro F1,test_macro_f1,True
cct,Caltech Camera Traps,SimpleShot + k-NN,"Non-empty detection uses SimpleShot, category recognition is cross-location k-NN.",non_empty_f1,Non-empty F1,k_100_cross_location_acc,False
cct,Caltech Camera Traps,SimpleShot + k-NN,"Non-empty detection uses SimpleShot, category recognition is cross-location k-NN.",k_10_cross_location_acc,k=10 cross-location accuracy,k_100_cross_location_acc,False
cct,Caltech Camera Traps,SimpleShot + k-NN,"Non-empty detection uses SimpleShot, category recognition is cross-location k-NN.",k_20_cross_location_acc,k=20 cross-location accuracy,k_100_cross_location_acc,False
cct,Caltech Camera Traps,SimpleShot + k-NN,"Non-empty detection uses SimpleShot, category recognition is cross-location k-NN.",k_100_cross_location_acc,k=100 cross-location accuracy,k_100_cross_location_acc,True
cub200,CUB-200-2011,Image retrieval,"Frozen embeddings are used for visual retrieval, mAP and Recall@K measure neighbor quality.",mean_average_precision,Mean average precision,mean_average_precision,True
cub200,CUB-200-2011,Image retrieval,"Frozen embeddings are used for visual retrieval, mAP and Recall@K measure neighbor quality.",r_at_1,Recall@1,mean_average_precision,False
cub200,CUB-200-2011,Image retrieval,"Frozen embeddings are used for visual retrieval, mAP and Recall@K measure neighbor quality.",r_at_5,Recall@5,mean_average_precision,False
cub200,CUB-200-2011,Image retrieval,"Frozen embeddings are used for visual retrieval, mAP and Recall@K measure neighbor quality.",r_at_10,Recall@10,mean_average_precision,False
fishnet,FishNet,MLP multi-label trait probe,Fish traits are predicted from frozen embeddings with a small MLP probe.,macro_f1,Macro F1,macro_f1,True
fishnet,FishNet,MLP multi-label trait probe,Fish traits are predicted from frozen embeddings with a small MLP probe.,exact_match_acc,Exact-match accuracy,macro_f1,False
flowers102,Flowers102,SimpleShot,Class prototypes are computed from frozen embeddings and evaluated on validation and test splits.,val_accuracy,Validation accuracy,test_accuracy,False
flowers102,Flowers102,SimpleShot,Class prototypes are computed from frozen embeddings and evaluated on validation and test splits.,test_accuracy,Test accuracy,test_accuracy,True
fungiclef,FungiCLEF2023,k-NN classification,Fungal species recognition over frozen embeddings with repeated k-NN runs.,k_1_acc,k=1 accuracy,k_3_acc,False
fungiclef,FungiCLEF2023,k-NN classification,Fungal species recognition over frozen embeddings with repeated k-NN runs.,k_3_acc,k=3 accuracy,k_3_acc,True
nabirds,NABirds,k-NN classification,Fine-grained bird classification from frozen embedding neighborhoods.,k_10_acc,k=10 accuracy,k_20_acc,False
nabirds,NABirds,k-NN classification,Fine-grained bird classification from frozen embedding neighborhoods.,k_20_acc,k=20 accuracy,k_20_acc,True
nabirds,NABirds,k-NN classification,Fine-grained bird classification from frozen embedding neighborhoods.,k_100_acc,k=100 accuracy,k_20_acc,False
newt,NeWT,SVM task suite,164 binary natural-world tasks are evaluated with SVMs over normalized frozen embeddings.,accuracy,Accuracy,accuracy,True
newt,NeWT,SVM task suite,164 binary natural-world tasks are evaluated with SVMs over normalized frozen embeddings.,cluster_appearance,Appearance cluster accuracy,accuracy,False
newt,NeWT,SVM task suite,164 binary natural-world tasks are evaluated with SVMs over normalized frozen embeddings.,cluster_behavior,Behavior cluster accuracy,accuracy,False
newt,NeWT,SVM task suite,164 binary natural-world tasks are evaluated with SVMs over normalized frozen embeddings.,cluster_context,Context cluster accuracy,accuracy,False
newt,NeWT,SVM task suite,164 binary natural-world tasks are evaluated with SVMs over normalized frozen embeddings.,cluster_counting,Counting cluster accuracy,accuracy,False
newt,NeWT,SVM task suite,164 binary natural-world tasks are evaluated with SVMs over normalized frozen embeddings.,cluster_gestalt,Gestalt cluster accuracy,accuracy,False
plantdoc,PlantDoc,Few-shot SimpleShot,Plant disease classification using 2-shot and 5-shot class prototypes.,2_shot_acc,2-shot accuracy,5_shot_acc,False
plantdoc,PlantDoc,Few-shot SimpleShot,Plant disease classification using 2-shot and 5-shot class prototypes.,5_shot_acc,5-shot accuracy,5_shot_acc,True
plantnet,PlantNet-300K,Few-shot SimpleShot,Plant species classification using 2-shot and 5-shot class prototypes.,2_shot_acc,2-shot accuracy,5_shot_acc,False
plantnet,PlantNet-300K,Few-shot SimpleShot,Plant species classification using 2-shot and 5-shot class prototypes.,5_shot_acc,5-shot accuracy,5_shot_acc,True
snakeclef,SnakeCLEF2023,Linear probe,Observation-level snake recognition with a trained linear head over frozen embeddings.,accuracy,Accuracy,accuracy,True
inat21-lp,iNaturalist 2021,Frozen backbone + linear probe,The image encoder is frozen and a 10k-class linear head is trained with light augmentation.,accuracy,Accuracy,accuracy,True
inat21-lp,iNaturalist 2021,Frozen backbone + linear probe,The image encoder is frozen and a 10k-class linear head is trained with light augmentation.,top_3_accuracy,Top-3 accuracy,accuracy,False
inat21-lp,iNaturalist 2021,Frozen backbone + linear probe,The image encoder is frozen and a 10k-class linear head is trained with light augmentation.,macro_f1,Macro F1,accuracy,False