{"record_type":"pith_number_record","schema_url":"https://pith.science/schemas/pith-number/v1.json","pith_number":"pith:2022:XUCPW5FBTH767BFSVERPKJNZMT","short_pith_number":"pith:XUCPW5FB","schema_version":"1.0","canonical_sha256":"bd04fb74a199ffef84b2a922f525b964d7c006dcfe5f34ce9b3262feae3367d1","source":{"kind":"arxiv","id":"2202.08005","version":3},"attestation_state":"computed","paper":{"title":"Should You Mask 15% in Masked Language Modeling?","license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","headline":"","cross_cats":["cs.LG"],"primary_cat":"cs.CL","authors_text":"Alexander Wettig, Danqi Chen, Tianyu Gao, Zexuan Zhong","submitted_at":"2022-02-16T11:42:34Z","abstract_excerpt":"Masked language models (MLMs) conventionally mask 15% of tokens due to the belief that more masking would leave insufficient context to learn good representations; this masking rate has been widely used, regardless of model sizes or masking strategies. In this work, we revisit this important choice of MLM pre-training. We first establish that 15% is not universally optimal, and larger models should adopt a higher masking rate. Specifically, we find that masking 40% outperforms 15% for BERT-large size models on GLUE and SQuAD. Interestingly, an extremely high masking rate of 80% can still prese"},"verification_status":{"content_addressed":true,"pith_receipt":true,"author_attested":false,"weak_author_claims":0,"strong_author_claims":0,"externally_anchored":false,"storage_verified":false,"citation_signatures":0,"replication_records":0,"graph_snapshot":true,"references_resolved":false,"formal_links_present":false},"canonical_record":{"source":{"id":"2202.08005","kind":"arxiv","version":3},"metadata":{"license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","primary_cat":"cs.CL","submitted_at":"2022-02-16T11:42:34Z","cross_cats_sorted":["cs.LG"],"title_canon_sha256":"97fff810efab9ef6eb1704ec0b386efc45d213e0887db251529e3880ec241952","abstract_canon_sha256":"f9fe0fe6d7cc2103022d4a634ccf54ed67f1aacf40917ca06b4c850164eee961"},"schema_version":"1.0"},"receipt":{"kind":"pith_receipt","key_id":"pith-v1-2026-05","algorithm":"ed25519","signed_at":"2026-07-05T05:40:30.455051Z","signature_b64":"64UpyD2GRwnxDeIL9lErNJD0cHf2a1dUjNPthqOkLPeEHjy+Etvozua1po8CDPksT1VLmFE8Q92p4CEu6upMDA==","signed_message":"canonical_sha256_bytes","builder_version":"pith-number-builder-2026-05-17-v1","receipt_version":"0.3","canonical_sha256":"bd04fb74a199ffef84b2a922f525b964d7c006dcfe5f34ce9b3262feae3367d1","last_reissued_at":"2026-07-05T05:40:30.454566Z","signature_status":"signed_v1","first_computed_at":"2026-07-05T05:40:30.454566Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"graph_snapshot":{"paper":{"title":"Should You Mask 15% in Masked Language Modeling?","license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","headline":"","cross_cats":["cs.LG"],"primary_cat":"cs.CL","authors_text":"Alexander Wettig, Danqi Chen, Tianyu Gao, Zexuan Zhong","submitted_at":"2022-02-16T11:42:34Z","abstract_excerpt":"Masked language models (MLMs) conventionally mask 15% of tokens due to the belief that more masking would leave insufficient context to learn good representations; this masking rate has been widely used, regardless of model sizes or masking strategies. In this work, we revisit this important choice of MLM pre-training. We first establish that 15% is not universally optimal, and larger models should adopt a higher masking rate. Specifically, we find that masking 40% outperforms 15% for BERT-large size models on GLUE and SQuAD. Interestingly, an extremely high masking rate of 80% can still prese"},"claims":{"count":0,"items":[],"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"source":{"id":"2202.08005","kind":"arxiv","version":3},"verdict":{"id":null,"model_set":{},"created_at":null,"strongest_claim":"","one_line_summary":"","pipeline_version":null,"weakest_assumption":"","pith_extraction_headline":""},"integrity":{"clean":true,"summary":{"advisory":0,"critical":0,"by_detector":{},"informational":0},"endpoint":"/pith/2202.08005/integrity.json","findings":[],"available":true,"detectors_run":[],"snapshot_sha256":"c28c3603d3b5d939e8dc4c7e95fa8dfce3d595e45f758748cecf8e644a296938"},"references":{"count":0,"sample":[],"resolved_work":0,"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57","internal_anchors":0},"formal_canon":{"evidence_count":0,"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"author_claims":{"count":0,"strong_count":0,"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"builder_version":"pith-number-builder-2026-05-17-v1"},"aliases":[{"alias_kind":"arxiv","alias_value":"2202.08005","created_at":"2026-07-05T05:40:30.454639+00:00"},{"alias_kind":"arxiv_version","alias_value":"2202.08005v3","created_at":"2026-07-05T05:40:30.454639+00:00"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2202.08005","created_at":"2026-07-05T05:40:30.454639+00:00"},{"alias_kind":"pith_short_12","alias_value":"XUCPW5FBTH76","created_at":"2026-07-05T05:40:30.454639+00:00"},{"alias_kind":"pith_short_16","alias_value":"XUCPW5FBTH767BFS","created_at":"2026-07-05T05:40:30.454639+00:00"},{"alias_kind":"pith_short_8","alias_value":"XUCPW5FB","created_at":"2026-07-05T05:40:30.454639+00:00"}],"events":[],"event_summary":{},"paper_claims":[],"inbound_citations":{"count":2,"internal_anchor_count":0,"sample":[{"citing_arxiv_id":"2412.13663","citing_title":"Smarter, Better, Faster, Longer: A Modern Bidirectional Encoder for Fast, Memory Efficient, and Long Context Finetuning and Inference","ref_index":198,"is_internal_anchor":false},{"citing_arxiv_id":"2605.19568","citing_title":"m3BERT: A Modern, Multi-lingual, Matryoshka Bidirectional Encoder","ref_index":46,"is_internal_anchor":false}]},"formal_canon":{"evidence_count":0,"sample":[],"anchors":[]},"links":{"html":"https://pith.science/pith/XUCPW5FBTH767BFSVERPKJNZMT","json":"https://pith.science/pith/XUCPW5FBTH767BFSVERPKJNZMT.json","graph_json":"https://pith.science/api/pith-number/XUCPW5FBTH767BFSVERPKJNZMT/graph.json","events_json":"https://pith.science/api/pith-number/XUCPW5FBTH767BFSVERPKJNZMT/events.json","paper":"https://pith.science/paper/XUCPW5FB"},"agent_actions":{"view_html":"https://pith.science/pith/XUCPW5FBTH767BFSVERPKJNZMT","download_json":"https://pith.science/pith/XUCPW5FBTH767BFSVERPKJNZMT.json","view_paper":"https://pith.science/paper/XUCPW5FB","resolve_alias":"https://pith.science/api/pith-number/resolve?arxiv=2202.08005&json=true","fetch_graph":"https://pith.science/api/pith-number/XUCPW5FBTH767BFSVERPKJNZMT/graph.json","fetch_events":"https://pith.science/api/pith-number/XUCPW5FBTH767BFSVERPKJNZMT/events.json","actions":{"anchor_timestamp":"https://pith.science/pith/XUCPW5FBTH767BFSVERPKJNZMT/action/timestamp_anchor","attest_storage":"https://pith.science/pith/XUCPW5FBTH767BFSVERPKJNZMT/action/storage_attestation","attest_author":"https://pith.science/pith/XUCPW5FBTH767BFSVERPKJNZMT/action/author_attestation","sign_citation":"https://pith.science/pith/XUCPW5FBTH767BFSVERPKJNZMT/action/citation_signature","submit_replication":"https://pith.science/pith/XUCPW5FBTH767BFSVERPKJNZMT/action/replication_record"}},"created_at":"2026-07-05T05:40:30.454639+00:00","updated_at":"2026-07-05T05:40:30.454639+00:00"}