{"id":"W4389616800","doi":"10.4995/jclr.2023.20408","title":"Autosupervisión de Alucinaciones en Grandes Modelos del Lenguaje: LLteaM","year":2023,"lang":"es","type":"article","venue":"Journal of Computer-Assisted Linguistic Research","topic":"Natural Language Processing Techniques","field":"Computer Science","cited_by":3,"is_retracted":false,"has_abstract":true,"ca_institutions":"WSP (Canada)","funders":"","keywords":"Humanities; Philosophy; Art","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":"metacan-v3-hybrid-931329e0061c","candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.003205477,0.0009885352,0.0007299766,0.001361892,0.001018654,0.00504533,0.001646444,0.001627145,0.006871531],"category_scores_gemma":[0.01224283,0.0006561161,0.001293164,0.0009047445,0.002032383,0.005649358,0.0028301,0.002074117,0.001576039],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.002431843,"about_ca_system_score_gemma":0.002825884,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.009612366,"about_ca_topic_score_gemma":0.007991914,"domain_scores_codex":[0.9973311,0.0008942827,0.0001521189,0.0007670174,0.0006423027,0.0002132509],"domain_scores_gemma":[0.9932679,0.003012324,0.0006151098,0.001471325,0.001308695,0.0003245581],"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":[0.00114713,0.0006103989,0.02954149,0.001128694,0.0003124598,0.001241251,0.009227505,0.1994884,0.02515763,0.2696634,0.009371306,0.4531104],"study_design_scores_gemma":[0.00007525766,0.0003469484,0.005691834,0.0002918804,0.0001709184,0.0004096516,0.001916212,0.7833371,0.01198353,0.1582491,0.03741949,0.000108033],"study_design_candidate":"bench_or_experimental","study_design_consensus":null,"genre_codex":"methods","genre_gemma":"empirical","genre_scores_codex":[0.2108765,0.001436975,0.7378401,0.002223045,0.0002227729,0.0003605747,0.0005039904,0.00522664,0.04130944],"genre_scores_gemma":[0.8319693,0.0007551806,0.1471292,0.0003566705,0.00006156675,0.0003445349,0.0007261718,0.0005442854,0.01811309],"genre_candidate":"empirical","genre_consensus":null,"teacher_disagreement_score":0.009612366,"threshold_uncertainty_score":0.02298754,"prediction_status":"machine_predicted_unvalidated"},"machine_scores":{"provisional":true,"baseline":true,"maturity_gate_passed":false,"score_opus":0.07493056431312235,"score_gpt":0.4078939249829552,"score_spread":0.3329633606698328,"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."}}