{"id":"W4415337949","doi":"10.48550/arxiv.2509.00404","title":"Metis: Training LLMs with FP4 Quantization","year":2025,"lang":"en","type":"preprint","venue":"ArXiv.org","topic":"Natural Language Processing Techniques","field":"Computer Science","cited_by":0,"is_retracted":false,"has_abstract":true,"ca_institutions":"","funders":"","keywords":"Metis; Quantization (signal processing); Singular value decomposition; Training (meteorology); Spectral line; Matching (statistics)","routes":{"ca_aff":false,"ca_fund":false,"ca_venue":false,"about_ca":true,"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":[0.00153837,0.001286525,0.0007666993,0.0006165592,0.000612879,0.001240183,0.002127001,0.001568292,0.01202752],"category_scores_gemma":[0.00731113,0.0005916813,0.0006692376,0.0006345619,0.0008943165,0.002090497,0.001935882,0.00272083,0.00544429],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.001316375,"about_ca_system_score_gemma":0.001812604,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.008994213,"about_ca_topic_score_gemma":0.0156984,"domain_scores_codex":[0.9990517,0.0002533541,0.00006434155,0.0002446068,0.0002764174,0.0001095307],"domain_scores_gemma":[0.9988292,0.0005389897,0.00005511609,0.0002549967,0.0002509322,0.00007084697],"domain_codex":null,"domain_gemma":null,"domain_candidate":null,"domain_consensus":null,"study_design_codex":"design_other","study_design_gemma":"not_applicable","study_design_scores_codex":[0.0007167323,0.0002149313,0.001494149,0.0003064164,0.0001280885,0.0002208473,0.0002806793,0.2903922,0.02307767,0.0214233,0.05170727,0.6100376],"study_design_scores_gemma":[0.00004066843,0.00005215703,0.0001285351,0.00002105127,0.000007131464,0.00003365218,0.00003949914,0.9774774,0.007813441,0.01030635,0.004067077,0.00001314032],"study_design_candidate":"not_applicable","study_design_consensus":null,"genre_codex":"methods","genre_gemma":"methods","genre_scores_codex":[0.03536779,0.0008157433,0.9199933,0.0008078956,0.0004389833,0.0001619567,0.001005414,0.03292511,0.008483792],"genre_scores_gemma":[0.3777367,0.0002531663,0.5994915,0.001117344,0.0001245609,0.0004670981,0.004307846,0.003735164,0.01276669],"genre_candidate":"methods","genre_consensus":"methods","teacher_disagreement_score":0.01202752,"threshold_uncertainty_score":0.040236,"prediction_status":"machine_predicted_unvalidated"},"machine_scores":{"provisional":true,"baseline":true,"maturity_gate_passed":false,"score_opus":0.05154907962914845,"score_gpt":0.3046366567186559,"score_spread":0.2530875770895075,"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."}}