{"id":"W4399971973","doi":"10.29012/jpc.880","title":"Differentially Private Fine-tuning of Language Models","year":2024,"lang":"en","type":"article","venue":"Journal of Privacy and Confidentiality","topic":"Privacy-Preserving Technologies in Data","field":"Computer Science","cited_by":36,"is_retracted":false,"has_abstract":true,"ca_institutions":"University of Waterloo","funders":"Natural Sciences and Engineering Research Council of Canada","keywords":"Computer science; Private speech; Base (topology); Artificial intelligence; Private information retrieval; Language model; Machine learning; Scale (ratio); Natural language processing; Data mining; Computer security; Mathematics","routes":{"ca_aff":true,"ca_fund":true,"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":["open_science"],"consensus_categories":["open_science"],"category_scores_codex":[0.001019289,0.0001469056,0.0003377169,0.000243094,0.00005655174,0.0003381638,0.009032965,0.00009929719,0.00002594536],"category_scores_gemma":[0.001909685,0.0001205374,0.0001398228,0.0002940085,0.0001221672,0.001451198,0.01741751,0.0004085863,0.000002123071],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.00002726185,"about_ca_system_score_gemma":0.0001058552,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.00006834903,"about_ca_topic_score_gemma":0.00000582421,"domain_scores_codex":[0.9982886,0.0001137567,0.0006813238,0.0002607289,0.000447256,0.0002083594],"domain_scores_gemma":[0.9971359,0.0001769845,0.0003536826,0.002130597,0.0001241609,0.00007867924],"domain_codex":null,"domain_gemma":null,"domain_candidate":null,"domain_consensus":null,"study_design_codex":"design_other","study_design_gemma":"theoretical_or_conceptual","study_design_scores_codex":[0.0001508583,0.0004176338,0.001460527,0.001721019,0.001248981,0.0008844934,0.009039147,0.00009509202,0.2344498,0.3106993,0.03748623,0.402347],"study_design_scores_gemma":[0.0006294205,0.0001751246,0.002165915,0.000665157,0.0001005055,0.0002870079,0.0001538385,0.1423403,0.04596534,0.805797,0.001418248,0.0003021347],"study_design_candidate":"theoretical_or_conceptual","study_design_consensus":null,"genre_codex":"methods","genre_gemma":"empirical","genre_scores_codex":[0.4086806,0.002038512,0.5856611,0.002819119,0.0005122567,0.00005380454,0.000008476463,0.00009061039,0.0001355013],"genre_scores_gemma":[0.9479591,0.0002429634,0.05164694,0.00003064017,0.00008057094,9.678644e-7,0.000001334448,0.000008387378,0.00002906315],"genre_candidate":"empirical","genre_consensus":null,"teacher_disagreement_score":0.5392784,"threshold_uncertainty_score":0.9963287,"prediction_status":"machine_predicted_unvalidated"},"machine_scores":{"provisional":true,"baseline":true,"maturity_gate_passed":false,"score_opus":0.03058623113808286,"score_gpt":0.2914109549588623,"score_spread":0.2608247238207794,"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."}}