{"id":"W4378770632","doi":"10.48550/arxiv.2305.17446","title":"Fine-tuning Happens in Tiny Subspaces: Exploring Intrinsic Task-specific Subspaces of Pre-trained Language Models","year":2023,"lang":"en","type":"preprint","venue":"arXiv (Cornell University)","topic":"Topic Modeling","field":"Computer Science","cited_by":0,"is_retracted":false,"has_abstract":true,"ca_institutions":"","funders":"Fundamental Research Funds for the Central Universities; Natural Sciences and Engineering Research Council of Canada; National Natural Science Foundation of China; Canadian Institute for Advanced Research","keywords":"Subspace topology; Linear subspace; Parameterized complexity; Computer science; Task (project management); Redundancy (engineering); Outlier; Perspective (graphical); Artificial intelligence; Process (computing); Machine learning; Algorithm; Mathematics","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":"metacan-v3-hybrid-931329e0061c","candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.001428205,0.0009652126,0.0008713803,0.0003724846,0.0003874019,0.001059729,0.0008011314,0.0008083454,0.0008511714],"category_scores_gemma":[0.01011131,0.0005466137,0.0007234687,0.0004075081,0.0008974843,0.002187913,0.001363953,0.002315358,0.000512017],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.0004530397,"about_ca_system_score_gemma":0.0008101074,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.002497102,"about_ca_topic_score_gemma":0.00303384,"domain_scores_codex":[0.9993991,0.0002265175,0.0000256145,0.0002081717,0.00006555994,0.00007515235],"domain_scores_gemma":[0.9975585,0.001454059,0.0001912942,0.0005569826,0.0001359237,0.0001031682],"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":[0.0002969086,0.0001734107,0.007452612,0.0001679565,0.0001846077,0.0001990273,0.0004852322,0.8238661,0.0313909,0.006597119,0.002016461,0.1271697],"study_design_scores_gemma":[0.0000100357,0.00005782233,0.0009153713,0.000009941091,0.00001402595,0.00003398898,0.00005253999,0.9867894,0.003624523,0.007958894,0.0005188195,0.00001463118],"study_design_candidate":"simulation_or_modeling","study_design_consensus":"simulation_or_modeling","genre_codex":"methods","genre_gemma":"empirical","genre_scores_codex":[0.3373802,0.0008302644,0.6576717,0.0005224346,0.00006171286,0.00005938271,0.0002095701,0.001460005,0.001804641],"genre_scores_gemma":[0.9329182,0.0002459832,0.06473338,0.0001838715,0.0000410416,0.0000722948,0.0005445647,0.0002858279,0.0009749257],"genre_candidate":"empirical","genre_consensus":null,"teacher_disagreement_score":0.002497102,"threshold_uncertainty_score":0.00755316,"prediction_status":"machine_predicted_unvalidated"},"machine_scores":{"provisional":true,"baseline":true,"maturity_gate_passed":false,"score_opus":0.1922405705722501,"score_gpt":0.2119595670151868,"score_spread":0.0197189964429367,"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."}}