{"id":"W4389518686","doi":"10.18653/v1/2023.emnlp-main.330","title":"Just Ask for Calibration: Strategies for Eliciting Calibrated Confidence Scores from Language Models Fine-Tuned with Human Feedback","year":2023,"lang":"en","type":"article","venue":"","topic":"Topic Modeling","field":"Computer Science","cited_by":83,"is_retracted":false,"has_abstract":true,"ca_institutions":"","funders":"Office of Naval Research; Canadian Institute for Advanced Research; National Science Foundation","keywords":"Ask price; Calibration; Computer science; Tian; Artificial intelligence; Statistics; Art; Mathematics; Literature","routes":{"ca_aff":false,"ca_fund":true,"ca_venue":false,"about_ca":false,"invisible_to_affiliation_only":true},"retraction":null,"screen":null,"direct_labels":[],"prediction":{"model_version":"metacan-v3-hybrid-931329e0061c","candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.0219118,0.002538387,0.001489459,0.001791164,0.0007326364,0.003026892,0.003561236,0.003567116,0.006749119],"category_scores_gemma":[0.2273429,0.001168809,0.0007423374,0.001454462,0.001070127,0.005614771,0.004765255,0.00361916,0.002998236],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.0009548627,"about_ca_system_score_gemma":0.001226966,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.001216832,"about_ca_topic_score_gemma":0.002218968,"domain_scores_codex":[0.9822886,0.01208872,0.0008493084,0.00258901,0.001767213,0.0004171777],"domain_scores_gemma":[0.8350257,0.1332784,0.005753972,0.01552753,0.008819737,0.001594655],"domain_codex":null,"domain_gemma":null,"domain_candidate":null,"domain_consensus":null,"study_design_codex":"design_other","study_design_gemma":"bench_or_experimental","study_design_scores_codex":[0.00372541,0.001656322,0.01894148,0.0008958505,0.0006517413,0.0004654553,0.004837494,0.06739479,0.03308478,0.01319668,0.02357158,0.8315784],"study_design_scores_gemma":[0.0007418176,0.0007997878,0.006252816,0.0001953094,0.0001918367,0.0002519766,0.001310944,0.8774102,0.03308703,0.07301553,0.006429397,0.0003133018],"study_design_candidate":"bench_or_experimental","study_design_consensus":null,"genre_codex":"methods","genre_gemma":"empirical","genre_scores_codex":[0.08019952,0.0005946994,0.9032762,0.001269607,0.0002008854,0.0006083102,0.0006834388,0.00881356,0.004353792],"genre_scores_gemma":[0.6746309,0.000151152,0.3205788,0.00065893,0.00009000453,0.0009388414,0.0009441965,0.0007159815,0.001291277],"genre_candidate":"empirical","genre_consensus":null,"teacher_disagreement_score":0.0219118,"threshold_uncertainty_score":0.115882,"prediction_status":"machine_predicted_unvalidated"},"machine_scores":{"provisional":true,"baseline":true,"maturity_gate_passed":false,"score_opus":0.07316623696227058,"score_gpt":0.3050333361367429,"score_spread":0.2318670991744723,"validation_status":"score_only:v0-immature-baseline","note":"Baseline scores from an immature model (maturity gate not passed). Scores rank; they never assert a category."}}