{"id":"W4403963877","doi":"10.48550/arxiv.2410.04350","title":"TIS-DPO: Token-level Importance Sampling for Direct Preference Optimization With Estimated Weights","year":2024,"lang":"en","type":"preprint","venue":"arXiv (Cornell University)","topic":"Data Management and Algorithms","field":"Computer Science","cited_by":0,"is_retracted":false,"has_abstract":true,"ca_institutions":"","funders":"Institute for Catastrophic Loss Reduction","keywords":"Preference; Security token; Sampling (signal processing); Statistics; Computer science; Mathematics; Computer network; Telecommunications","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":[0.0002043484,0.0003639228,0.0003150377,0.0002821886,0.0001746082,0.0004731712,0.001727939,0.0001523461,0.00001914699],"category_scores_gemma":[0.00001757904,0.0003534627,0.0001060346,0.0007057415,0.00006072815,0.000687737,0.002060961,0.0003162043,0.00003326595],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.000156366,"about_ca_system_score_gemma":0.0001709515,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.00005215238,"about_ca_topic_score_gemma":0.00004098823,"domain_scores_codex":[0.9978235,0.00003342846,0.0002077453,0.00143918,0.0001211272,0.0003749641],"domain_scores_gemma":[0.9983155,0.00008846908,0.000220881,0.001085569,0.0001746143,0.0001149827],"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":[0.0000372121,0.00006111454,0.0005805784,0.0003059218,0.0001874927,0.0001325302,0.00008930836,0.8755519,0.000002809067,0.1199006,0.0007203282,0.002430238],"study_design_scores_gemma":[0.0003475387,0.0000620206,0.0003455693,0.0003450976,0.0001359883,0.000001940918,0.00001148322,0.97495,0.00005264351,0.02274328,0.0005131065,0.0004913952],"study_design_candidate":"simulation_or_modeling","study_design_consensus":"simulation_or_modeling","genre_codex":"methods","genre_gemma":"empirical","genre_scores_codex":[0.009287802,0.000101179,0.9853404,0.0000827534,0.0005752729,0.0006803091,0.0001811545,0.0005590459,0.003192068],"genre_scores_gemma":[0.6300984,0.0002714503,0.3640864,0.00006341533,0.0001092113,0.00001247171,0.0004924071,0.00004684027,0.004819468],"genre_candidate":"methods","genre_consensus":null,"teacher_disagreement_score":0.6212541,"threshold_uncertainty_score":0.9998918,"prediction_status":"machine_predicted_unvalidated"},"machine_scores":{"provisional":true,"baseline":true,"maturity_gate_passed":false,"score_opus":0.2090452637902154,"score_gpt":0.2221080567722163,"score_spread":0.01306279298200086,"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."}}