{"id":"W4225817241","doi":"10.1007/s10489-022-03944-z","title":"Reward modeling for mitigating toxicity in transformer-based language models","year":2022,"lang":"en","type":"article","venue":"Applied Intelligence","topic":"Hate Speech and Cyberbullying Detection","field":"Computer Science","cited_by":0,"is_retracted":false,"has_abstract":false,"ca_institutions":"Concordia University","funders":"","keywords":"Computer science; Language model; Transformer; Artificial intelligence; Unintended consequences; Detoxification (alternative medicine); Machine learning; Natural language processing; Engineering","routes":{"ca_aff":true,"ca_fund":false,"ca_venue":false,"about_ca":false,"invisible_to_affiliation_only":false},"retraction":null,"screen":null,"direct_labels":[],"prediction":{"model_version":"codex-gemma-dda1882f352a","candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.0005814824,0.0001311698,0.0001473179,0.000140511,0.0002710056,0.00005726892,0.0006581361,0.00004028674,0.00001904981],"category_scores_gemma":[0.00001441415,0.0001477268,0.00007204871,0.0005148761,0.000016797,0.0001763995,0.00006345275,0.0002528191,0.000006821844],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.0001187904,"about_ca_system_score_gemma":0.00008191845,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.0001185478,"about_ca_topic_score_gemma":0.0000374474,"domain_scores_codex":[0.9986038,0.00003138212,0.0002993515,0.0004307163,0.0002728953,0.0003618908],"domain_scores_gemma":[0.9994659,0.00008674133,0.00005242334,0.0003098497,0.00002565328,0.00005941341],"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":[0.00002710817,0.00004792222,0.000002159892,0.00001845108,0.000003180857,0.000003631855,0.002206283,0.8398613,0.01534975,0.03236042,0.000008660115,0.1101112],"study_design_scores_gemma":[0.0001101102,0.00006841567,3.205515e-7,0.000008156614,0.000001802903,0.000002665364,0.0008896684,0.833937,0.1444487,0.02032353,0.00005555093,0.0001540782],"study_design_candidate":"simulation_or_modeling","study_design_consensus":"simulation_or_modeling","genre_codex":"methods","genre_gemma":"empirical","genre_scores_codex":[0.1178526,0.00004956374,0.8795109,0.0001536685,0.0001109863,0.0004269107,0.000005727672,0.0001483927,0.001741251],"genre_scores_gemma":[0.9316757,0.000003524703,0.06739972,0.0004168903,0.00002232561,0.0004380691,0.000006485963,0.00001298508,0.00002425267],"genre_candidate":"empirical","genre_consensus":null,"teacher_disagreement_score":0.8138232,"threshold_uncertainty_score":0.6024124,"prediction_status":"machine_predicted_unvalidated"},"machine_scores":{"provisional":true,"baseline":true,"maturity_gate_passed":false,"score_opus":0.0284011211217057,"score_gpt":0.2561019994612627,"score_spread":0.227700878339557,"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."}}