{"record_type":"pith_number_record","schema_url":"https://pith.science/schemas/pith-number/v1.json","pith_number":"pith:2019:542YWXACH7F6UJBTE74HXHEDG3","short_pith_number":"pith:542YWXAC","schema_version":"1.0","canonical_sha256":"ef358b5c023fcbea243327f87b9c8336e09b36a828dedb27eaba6f8c6b1e1dea","source":{"kind":"arxiv","id":"1908.05620","version":1},"attestation_state":"computed","paper":{"title":"Visualizing and Understanding the Effectiveness of BERT","license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","headline":"","cross_cats":["cs.LG"],"primary_cat":"cs.CL","authors_text":"Furu Wei, Ke Xu, Li Dong, Yaru Hao","submitted_at":"2019-08-15T16:11:45Z","abstract_excerpt":"Language model pre-training, such as BERT, has achieved remarkable results in many NLP tasks. However, it is unclear why the pre-training-then-fine-tuning paradigm can improve performance and generalization capability across different tasks. In this paper, we propose to visualize loss landscapes and optimization trajectories of fine-tuning BERT on specific datasets. First, we find that pre-training reaches a good initial point across downstream tasks, which leads to wider optima and easier optimization compared with training from scratch. We also demonstrate that the fine-tuning procedure is r"},"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":"1908.05620","kind":"arxiv","version":1},"metadata":{"license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","primary_cat":"cs.CL","submitted_at":"2019-08-15T16:11:45Z","cross_cats_sorted":["cs.LG"],"title_canon_sha256":"5ad4d22ba05c741c18696b50f33ac97fcaa5aecffce749c4c9d9238b1f2a8266","abstract_canon_sha256":"23670d85a45606e9c4691dc61698c0401422b16781cfc0ed485e26e6d7999186"},"schema_version":"1.0"},"receipt":{"kind":"pith_receipt","key_id":"pith-v1-2026-05","algorithm":"ed25519","signed_at":"2026-07-04T23:56:54.377829Z","signature_b64":"lswrdIKzqgM5r/jIngjhjdS8D7eCoYd7lIO30+0iqxeZGjjyZV/XAF5ncy5FC5pZ/rXxlaTKwjTcx7UpjBWCCA==","signed_message":"canonical_sha256_bytes","builder_version":"pith-number-builder-2026-05-17-v1","receipt_version":"0.3","canonical_sha256":"ef358b5c023fcbea243327f87b9c8336e09b36a828dedb27eaba6f8c6b1e1dea","last_reissued_at":"2026-07-04T23:56:54.377415Z","signature_status":"signed_v1","first_computed_at":"2026-07-04T23:56:54.377415Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"graph_snapshot":{"paper":{"title":"Visualizing and Understanding the Effectiveness of BERT","license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","headline":"","cross_cats":["cs.LG"],"primary_cat":"cs.CL","authors_text":"Furu Wei, Ke Xu, Li Dong, Yaru Hao","submitted_at":"2019-08-15T16:11:45Z","abstract_excerpt":"Language model pre-training, such as BERT, has achieved remarkable results in many NLP tasks. However, it is unclear why the pre-training-then-fine-tuning paradigm can improve performance and generalization capability across different tasks. In this paper, we propose to visualize loss landscapes and optimization trajectories of fine-tuning BERT on specific datasets. First, we find that pre-training reaches a good initial point across downstream tasks, which leads to wider optima and easier optimization compared with training from scratch. We also demonstrate that the fine-tuning procedure is r"},"claims":{"count":0,"items":[],"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"source":{"id":"1908.05620","kind":"arxiv","version":1},"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/1908.05620/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":"1908.05620","created_at":"2026-07-04T23:56:54.377472+00:00"},{"alias_kind":"arxiv_version","alias_value":"1908.05620v1","created_at":"2026-07-04T23:56:54.377472+00:00"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.1908.05620","created_at":"2026-07-04T23:56:54.377472+00:00"},{"alias_kind":"pith_short_12","alias_value":"542YWXACH7F6","created_at":"2026-07-04T23:56:54.377472+00:00"},{"alias_kind":"pith_short_16","alias_value":"542YWXACH7F6UJBT","created_at":"2026-07-04T23:56:54.377472+00:00"},{"alias_kind":"pith_short_8","alias_value":"542YWXAC","created_at":"2026-07-04T23:56:54.377472+00:00"}],"events":[],"event_summary":{},"paper_claims":[],"inbound_citations":{"count":1,"internal_anchor_count":1,"sample":[{"citing_arxiv_id":"2508.17233","citing_title":"Module-Aware Parameter-Efficient Machine Unlearning on Transformers","ref_index":14,"is_internal_anchor":true}]},"formal_canon":{"evidence_count":0,"sample":[],"anchors":[]},"links":{"html":"https://pith.science/pith/542YWXACH7F6UJBTE74HXHEDG3","json":"https://pith.science/pith/542YWXACH7F6UJBTE74HXHEDG3.json","graph_json":"https://pith.science/api/pith-number/542YWXACH7F6UJBTE74HXHEDG3/graph.json","events_json":"https://pith.science/api/pith-number/542YWXACH7F6UJBTE74HXHEDG3/events.json","paper":"https://pith.science/paper/542YWXAC"},"agent_actions":{"view_html":"https://pith.science/pith/542YWXACH7F6UJBTE74HXHEDG3","download_json":"https://pith.science/pith/542YWXACH7F6UJBTE74HXHEDG3.json","view_paper":"https://pith.science/paper/542YWXAC","resolve_alias":"https://pith.science/api/pith-number/resolve?arxiv=1908.05620&json=true","fetch_graph":"https://pith.science/api/pith-number/542YWXACH7F6UJBTE74HXHEDG3/graph.json","fetch_events":"https://pith.science/api/pith-number/542YWXACH7F6UJBTE74HXHEDG3/events.json","actions":{"anchor_timestamp":"https://pith.science/pith/542YWXACH7F6UJBTE74HXHEDG3/action/timestamp_anchor","attest_storage":"https://pith.science/pith/542YWXACH7F6UJBTE74HXHEDG3/action/storage_attestation","attest_author":"https://pith.science/pith/542YWXACH7F6UJBTE74HXHEDG3/action/author_attestation","sign_citation":"https://pith.science/pith/542YWXACH7F6UJBTE74HXHEDG3/action/citation_signature","submit_replication":"https://pith.science/pith/542YWXACH7F6UJBTE74HXHEDG3/action/replication_record"}},"created_at":"2026-07-04T23:56:54.377472+00:00","updated_at":"2026-07-04T23:56:54.377472+00:00"}