{"id":"W4378513156","doi":"10.48550/arxiv.2305.14975","title":"Just Ask for Calibration: Strategies for Eliciting Calibrated Confidence Scores from Language Models Fine-Tuned with Human Feedback","year":2023,"lang":"en","type":"preprint","venue":"arXiv (Cornell University)","topic":"Topic Modeling","field":"Computer Science","cited_by":5,"is_retracted":false,"has_abstract":true,"ca_institutions":"","funders":"Office of Naval Research; Canadian Institute for Advanced Research; National Science Foundation","keywords":"Calibration; Computer science; Ask price; Confidence interval; Trustworthiness; Machine learning; Artificial intelligence; Deferral; Statistics; Mathematics","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":[0.01462665,0.002438904,0.001134942,0.001781933,0.0005735882,0.002501918,0.002963699,0.002699355,0.003700485],"category_scores_gemma":[0.1493353,0.0008944247,0.0007105935,0.0009782219,0.00126143,0.004762216,0.003143066,0.004215914,0.001803382],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.001152131,"about_ca_system_score_gemma":0.001388631,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.0021137,"about_ca_topic_score_gemma":0.003167705,"domain_scores_codex":[0.989767,0.005933,0.0005437377,0.002029993,0.001377279,0.0003489662],"domain_scores_gemma":[0.901718,0.07649145,0.004771955,0.01024044,0.005597527,0.001180598],"domain_codex":null,"domain_gemma":null,"domain_candidate":null,"domain_consensus":null,"study_design_codex":"design_other","study_design_gemma":"simulation_or_modeling","study_design_scores_codex":[0.002490764,0.001140509,0.03824137,0.001029746,0.0006416044,0.0005121114,0.004558253,0.1930494,0.04369497,0.0135174,0.01945367,0.6816702],"study_design_scores_gemma":[0.0002287684,0.0003909277,0.00656185,0.0001149074,0.00008004279,0.0001830202,0.0004631108,0.930155,0.03140127,0.02665642,0.003587079,0.0001776228],"study_design_candidate":"simulation_or_modeling","study_design_consensus":null,"genre_codex":"methods","genre_gemma":"empirical","genre_scores_codex":[0.140399,0.0006037183,0.8306122,0.001207877,0.0001444087,0.0004664489,0.001055896,0.02142539,0.004085085],"genre_scores_gemma":[0.7737534,0.0001096125,0.2218184,0.0005027824,0.00006440459,0.000450341,0.001182337,0.0008664365,0.001252317],"genre_candidate":"empirical","genre_consensus":null,"teacher_disagreement_score":0.01462665,"threshold_uncertainty_score":0.07735407,"prediction_status":"machine_predicted_unvalidated"},"machine_scores":{"provisional":true,"baseline":true,"maturity_gate_passed":false,"score_opus":0.1723460406433187,"score_gpt":0.2401572498743322,"score_spread":0.06781120923101344,"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."}}