{"id":"W4398836342","doi":"10.48550/arxiv.2405.14577","title":"Representation Noising: A Defence Mechanism Against Harmful Finetuning","year":2024,"lang":"en","type":"preprint","venue":"arXiv (Cornell University)","topic":"Scientific Computing and Data Management","field":"Decision Sciences","cited_by":0,"is_retracted":false,"has_abstract":true,"ca_institutions":"","funders":"Natural Sciences and Engineering Research Council of Canada; Research Nova Scotia; Alliance de recherche numérique du Canada; Killam Trusts; Bundesministerium für Bildung und Forschung; Canadian Institute for Advanced Research","keywords":"Representation (politics); Fine-tuning; Computer science; Artificial intelligence; Political science; Physics; Law; Politics","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","scholarly_communication","open_science","insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.003270036,0.0003350012,0.0004141725,0.001039668,0.0003154327,0.001397778,0.002672889,0.0001927889,0.0002173931],"category_scores_gemma":[0.001376739,0.0003307422,0.0003545739,0.002558395,0.000137522,0.0002818359,0.008296098,0.0006834401,0.001968894],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.0001929238,"about_ca_system_score_gemma":0.0001761707,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.0001752918,"about_ca_topic_score_gemma":0.00006292021,"domain_scores_codex":[0.9951799,0.0003167271,0.0005602412,0.002876892,0.0006520281,0.0004142484],"domain_scores_gemma":[0.9955181,0.0006427519,0.00044586,0.002808162,0.0003805172,0.0002045356],"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":[0.00006129738,0.0001231726,0.001651991,0.000104849,0.000188348,0.001341036,0.00124089,0.671158,0.0002796902,0.233296,0.06546157,0.02509326],"study_design_scores_gemma":[0.0001519349,0.00001348902,0.0003373589,0.0001480304,0.00007785658,0.000002061111,0.0009419073,0.5957696,0.000109444,0.3986249,0.003504182,0.0003192234],"study_design_candidate":"simulation_or_modeling","study_design_consensus":"simulation_or_modeling","genre_codex":"empirical","genre_gemma":"empirical","genre_scores_codex":[0.6778721,0.00009051529,0.2957101,0.00047608,0.005071936,0.0003893804,0.0000785698,0.0004006057,0.01991074],"genre_scores_gemma":[0.9750903,0.00003868059,0.0008947282,0.0001465552,0.0001476108,9.454731e-7,0.00006768321,0.00002179625,0.02359165],"genre_candidate":"empirical","genre_consensus":"empirical","teacher_disagreement_score":0.2972183,"threshold_uncertainty_score":0.9999145,"prediction_status":"machine_predicted_unvalidated"},"machine_scores":{"provisional":true,"baseline":true,"maturity_gate_passed":false,"score_opus":0.3478307503312511,"score_gpt":0.2981831477082125,"score_spread":0.04964760262303858,"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."}}