{"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":"metacan-v3-hybrid-931329e0061c","candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.002885436,0.0007006958,0.000598417,0.0006742512,0.0006514197,0.001086202,0.0007765963,0.0007225134,0.003087199],"category_scores_gemma":[0.02160734,0.0003391703,0.0005094757,0.0008981344,0.001567356,0.003216935,0.001762989,0.002036686,0.0007167872],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.0005533749,"about_ca_system_score_gemma":0.0007691999,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.002960443,"about_ca_topic_score_gemma":0.004123493,"domain_scores_codex":[0.9988816,0.000333281,0.00009092283,0.0002953498,0.0002838014,0.0001150508],"domain_scores_gemma":[0.9882839,0.006957948,0.000657242,0.003237437,0.000666721,0.0001967509],"domain_codex":null,"domain_gemma":null,"domain_candidate":null,"domain_consensus":null,"study_design_codex":"simulation_or_modeling","study_design_gemma":"bench_or_experimental","study_design_scores_codex":[0.0006155979,0.0002288022,0.02191973,0.0004003076,0.0002492085,0.001276053,0.001628279,0.5559548,0.05280798,0.07770382,0.01153157,0.2756839],"study_design_scores_gemma":[0.00003885883,0.0001722813,0.008572039,0.00006179648,0.00005297037,0.0007058501,0.0004287443,0.8725764,0.02154285,0.08992098,0.005864578,0.00006272356],"study_design_candidate":"bench_or_experimental","study_design_consensus":null,"genre_codex":"empirical","genre_gemma":"empirical","genre_scores_codex":[0.5132662,0.001440057,0.4586875,0.004912005,0.0002226594,0.0001830785,0.001745793,0.003110001,0.01643264],"genre_scores_gemma":[0.9434876,0.0005647842,0.05113087,0.0004503269,0.0001712159,0.0001332142,0.00179313,0.0003038556,0.001965037],"genre_candidate":"empirical","genre_consensus":"empirical","teacher_disagreement_score":0.003087199,"threshold_uncertainty_score":0.0152598,"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."}}