{"datasets":[{"key":"cs10k","short":"CS10k","label":"ConstructionSite-10k","color":"#1b2a49","images":10013,"items":60031,"asia":0.956,"western":0.037},{"key":"chv","short":"CHV","label":"CHV","color":"#b8236b","images":1284,"items":3852,"asia":0.306,"western":0.676},{"key":"gdut","short":"GDUT","label":"GDUT-HWD","color":"#1f5f3f","images":3073,"items":6146,"asia":0.733,"western":0.259},{"key":"shwd","short":"SHWD","label":"SHWD","color":"#c9a227","images":3152,"items":6304,"asia":0.777,"western":0.214},{"key":"pictor","short":"Pictor","label":"Pictor-v3","color":"#1d6f7a","images":701,"items":1402,"asia":0.897,"western":0.1},{"key":"sh17","short":"SH17","label":"SH17","color":"#7a1f3d","images":1001,"items":3003,"asia":0.277,"western":0.68},{"key":"roadwork","short":"ROADWork","label":"ROADWork","color":"#9aa1ad","images":7416,"items":22228,"asia":0.0,"western":1.0}],"models":[{"key":"qwen25vl_7b","short":"Qwen-7B","label":"Qwen2.5-VL-7B","finetuned":false},{"key":"qwen25vl_3b","short":"Qwen-3B","label":"Qwen2.5-VL-3B","finetuned":false},{"key":"llava_ov_7b","short":"LLaVA-OV-7B","label":"LLaVA-OneVision-7B","finetuned":false},{"key":"internvl3_8b","short":"InternVL3-8B","label":"InternVL3-8B","finetuned":false},{"key":"qwen25vl_7b_ft","short":"Qwen-7B ft","label":"Qwen2.5-VL-7B + QLoRA","finetuned":true},{"key":"qwen25vl_3b_ft","short":"Qwen-3B ft","label":"Qwen2.5-VL-3B + QLoRA","finetuned":true}],"estimators":[{"code":"U1","key":"verbal","name":"verbalized confidence","extra_passes":0},{"code":"U2a","key":"tokenprob","name":"answer-token probability","extra_passes":0},{"code":"U2b","key":"margin","name":"top-two margin","extra_passes":0},{"code":"U2c","key":"neg_entropy","name":"negative answer entropy","extra_passes":0},{"code":"U3","key":"self_consistency","name":"self-consistency vote","extra_passes":10},{"code":"U4","key":"neg_vote_entropy","name":"predictive vote entropy","extra_passes":10},{"code":"U5","key":"probe","name":"hidden-state probe","extra_passes":0},{"code":"U6","key":"conformal","name":"conformal risk control","extra_passes":0},{"code":"U7","key":"ensemble","name":"cross-model ensemble","extra_passes":3}],"headline":{"items":102966,"images":26640,"datasets":7,"models":6,"token_auroc":0.495,"probe_auroc":0.818,"probe_layer":28,"crc_alpha":0.05,"crc_coverage":0.083115752829,"crc_fnr_auto":0.037991858887,"crc_fnr_none":0.298507462687,"shift_dataset":"shwd","shift_fnr":0.639380530973,"shift_factor":16.83,"records_screened":9694,"autcon_papers":4558}}