{"id":"W4385573538","doi":"10.18653/v1/2022.emnlp-main.619","title":"Intriguing Properties of Compression on Multilingual Models","year":2022,"lang":"en","type":"article","venue":"","topic":"Topic Modeling","field":"Computer Science","cited_by":4,"is_retracted":false,"has_abstract":true,"ca_institutions":"University of Waterloo","funders":"","keywords":"Computer science; Compression (physics); Robustness (evolution); Generalization; Multilingualism; Language model; Data compression; Data compression ratio; Scaling; Natural language processing; Artificial intelligence; Image compression; Linguistics; Mathematics","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":"codex-gemma-dda1882f352a","candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.0001686808,0.00005263004,0.00008324903,0.00006366533,0.0001129277,0.0000162362,0.0005262123,0.00001114924,0.00002056239],"category_scores_gemma":[0.00001093364,0.0000410497,0.00002609015,0.00009053245,0.00001026695,0.0001392306,0.0005923271,0.0001071618,0.000001798971],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.00002474094,"about_ca_system_score_gemma":0.00003074866,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.00007110418,"about_ca_topic_score_gemma":9.171796e-7,"domain_scores_codex":[0.9992195,0.0000503279,0.0001538371,0.0001888484,0.0002825173,0.0001049833],"domain_scores_gemma":[0.9995643,0.00001941575,0.00004359076,0.000322083,0.00002863418,0.00002201538],"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":[0.00001541341,0.00009752289,0.00002932392,0.00001532493,0.000004951185,0.000003418142,0.00352436,0.8350475,0.02014358,0.08583685,0.0000879838,0.05519373],"study_design_scores_gemma":[0.0001304525,0.0000428824,0.00000379236,0.0000139566,5.519868e-7,0.000001945436,0.0001226659,0.9462967,0.05157544,0.001600536,0.0001590162,0.00005209325],"study_design_candidate":"simulation_or_modeling","study_design_consensus":"simulation_or_modeling","genre_codex":"empirical","genre_gemma":"empirical","genre_scores_codex":[0.5235251,0.00007538135,0.4725234,0.0002344242,0.0002165489,0.00009504372,3.935205e-7,0.0001043121,0.003225465],"genre_scores_gemma":[0.9757121,8.456605e-7,0.0238002,0.0001625785,0.00001499924,0.000009307558,1.851392e-7,0.000003275223,0.0002965099],"genre_candidate":"empirical","genre_consensus":"empirical","teacher_disagreement_score":0.4521871,"threshold_uncertainty_score":0.1673958,"prediction_status":"machine_predicted_unvalidated"},"machine_scores":{"provisional":true,"baseline":true,"maturity_gate_passed":false,"score_opus":0.07661469801205291,"score_gpt":0.2583159456123127,"score_spread":0.1817012476002599,"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."}}