{"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":"codex-gemma-dda1882f352a","candidate_categories":["metaepi_narrow"],"consensus_categories":[],"category_scores_codex":[0.0002054947,0.0003985867,0.0004545238,0.000200102,0.0003325877,0.0005875294,0.001662113,0.0002881847,0.00001051291],"category_scores_gemma":[0.00002929666,0.0004255482,0.0001705088,0.0003654136,0.00009852007,0.001183684,0.0008021186,0.000326999,0.000005274554],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.0001105856,"about_ca_system_score_gemma":0.0004853399,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.002823259,"about_ca_topic_score_gemma":0.001883841,"domain_scores_codex":[0.9975364,0.0000779179,0.0003349859,0.001455159,0.0001359541,0.000459553],"domain_scores_gemma":[0.9977434,0.0003712078,0.0003571532,0.00112638,0.0002633578,0.0001385196],"domain_codex":null,"domain_gemma":null,"domain_candidate":null,"domain_consensus":null,"study_design_codex":"simulation_or_modeling","study_design_gemma":"simulation_or_modeling","study_design_scores_codex":[0.00004172054,0.00001726907,0.0001065213,0.000132862,0.00007989551,0.00005596975,0.001083392,0.6515366,0.0004169795,0.3463104,0.0001503375,0.00006800035],"study_design_scores_gemma":[0.0006972748,0.00005968539,0.00003740011,0.0002379742,0.00006363206,7.651056e-7,0.00109346,0.7704148,0.0006455069,0.22634,0.000006031558,0.0004034466],"study_design_candidate":"simulation_or_modeling","study_design_consensus":"simulation_or_modeling","genre_codex":"methods","genre_gemma":"empirical","genre_scores_codex":[0.2709736,0.00002970749,0.7270861,0.0001444462,0.0002119726,0.000742916,0.0001564482,0.0005101893,0.0001446411],"genre_scores_gemma":[0.9700966,0.000008083142,0.02828226,0.00009002468,0.0002155807,0.00002112794,0.0002521263,0.00005001976,0.0009841091],"genre_candidate":"empirical","genre_consensus":null,"teacher_disagreement_score":0.6991231,"threshold_uncertainty_score":0.9998196,"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."}}