{"record_type":"pith_number_record","schema_url":"https://pith.science/schemas/pith-number/v1.json","pith_number":"pith:2024:WDEIK2RSZEFHKXYVQLWPF4AYTN","short_pith_number":"pith:WDEIK2RS","schema_version":"1.0","canonical_sha256":"b0c8856a32c90a755f1582ecf2f0189b4250240ca7b7fafd29c53cbfeb5e0da1","source":{"kind":"arxiv","id":"2406.14833","version":2},"attestation_state":"computed","paper":{"title":"Efficient Continual Pre-training by Mitigating the Stability Gap","license":"http://creativecommons.org/licenses/by/4.0/","headline":"","cross_cats":[],"primary_cat":"cs.CL","authors_text":"Dongyan Zhao, Huishuai Zhang, Jie Fu, Yiduo Guo, Yikang Shen","submitted_at":"2024-06-21T02:28:37Z","abstract_excerpt":"Continual pre-training has increasingly become the predominant approach for adapting Large Language Models (LLMs) to new domains. This process involves updating the pre-trained LLM with a corpus from a new domain, resulting in a shift in the training distribution. To study the behavior of LLMs during this shift, we measured the model's performance throughout the continual pre-training process. we observed a temporary performance drop at the beginning, followed by a recovery phase, a phenomenon known as the \"stability gap,\" previously noted in vision models classifying new classes. To address t"},"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":"2406.14833","kind":"arxiv","version":2},"metadata":{"license":"http://creativecommons.org/licenses/by/4.0/","primary_cat":"cs.CL","submitted_at":"2024-06-21T02:28:37Z","cross_cats_sorted":[],"title_canon_sha256":"5c9e318de05afc1668938fe19fa52a4055088d3ef0d602664a88fc273960c672","abstract_canon_sha256":"2907793a62213c0c2d46fce0a8d736fb518342418dcc49aebb7b5cdb38d3a849"},"schema_version":"1.0"},"receipt":{"kind":"pith_receipt","key_id":"pith-v1-2026-05","algorithm":"ed25519","signed_at":"2026-07-05T08:37:20.937073Z","signature_b64":"vHwR0V40oj08MWSeIHBTQMDJdeOX+oiOyF0M5DEooul5ug7OK3gipr1Uy4A+3OBF0SPu6wEVu09fRTO1pLCrDQ==","signed_message":"canonical_sha256_bytes","builder_version":"pith-number-builder-2026-05-17-v1","receipt_version":"0.3","canonical_sha256":"b0c8856a32c90a755f1582ecf2f0189b4250240ca7b7fafd29c53cbfeb5e0da1","last_reissued_at":"2026-07-05T08:37:20.936635Z","signature_status":"signed_v1","first_computed_at":"2026-07-05T08:37:20.936635Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"graph_snapshot":{"paper":{"title":"Efficient Continual Pre-training by Mitigating the Stability Gap","license":"http://creativecommons.org/licenses/by/4.0/","headline":"","cross_cats":[],"primary_cat":"cs.CL","authors_text":"Dongyan Zhao, Huishuai Zhang, Jie Fu, Yiduo Guo, Yikang Shen","submitted_at":"2024-06-21T02:28:37Z","abstract_excerpt":"Continual pre-training has increasingly become the predominant approach for adapting Large Language Models (LLMs) to new domains. This process involves updating the pre-trained LLM with a corpus from a new domain, resulting in a shift in the training distribution. To study the behavior of LLMs during this shift, we measured the model's performance throughout the continual pre-training process. we observed a temporary performance drop at the beginning, followed by a recovery phase, a phenomenon known as the \"stability gap,\" previously noted in vision models classifying new classes. To address t"},"claims":{"count":0,"items":[],"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"source":{"id":"2406.14833","kind":"arxiv","version":2},"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/2406.14833/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":"2406.14833","created_at":"2026-07-05T08:37:20.936692+00:00"},{"alias_kind":"arxiv_version","alias_value":"2406.14833v2","created_at":"2026-07-05T08:37:20.936692+00:00"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2406.14833","created_at":"2026-07-05T08:37:20.936692+00:00"},{"alias_kind":"pith_short_12","alias_value":"WDEIK2RSZEFH","created_at":"2026-07-05T08:37:20.936692+00:00"},{"alias_kind":"pith_short_16","alias_value":"WDEIK2RSZEFHKXYV","created_at":"2026-07-05T08:37:20.936692+00:00"},{"alias_kind":"pith_short_8","alias_value":"WDEIK2RS","created_at":"2026-07-05T08:37:20.936692+00:00"}],"events":[],"event_summary":{},"paper_claims":[],"inbound_citations":{"count":2,"internal_anchor_count":0,"sample":[{"citing_arxiv_id":"2606.24901","citing_title":"LLM Evolution as an Industry-Scale Ecosystem: A Lifecycle Perspective on Continual Learning","ref_index":34,"is_internal_anchor":false},{"citing_arxiv_id":"2409.04777","citing_title":"Optimization Hyper-parameter Laws for Large Language Models","ref_index":12,"is_internal_anchor":false}]},"formal_canon":{"evidence_count":0,"sample":[],"anchors":[]},"links":{"html":"https://pith.science/pith/WDEIK2RSZEFHKXYVQLWPF4AYTN","json":"https://pith.science/pith/WDEIK2RSZEFHKXYVQLWPF4AYTN.json","graph_json":"https://pith.science/api/pith-number/WDEIK2RSZEFHKXYVQLWPF4AYTN/graph.json","events_json":"https://pith.science/api/pith-number/WDEIK2RSZEFHKXYVQLWPF4AYTN/events.json","paper":"https://pith.science/paper/WDEIK2RS"},"agent_actions":{"view_html":"https://pith.science/pith/WDEIK2RSZEFHKXYVQLWPF4AYTN","download_json":"https://pith.science/pith/WDEIK2RSZEFHKXYVQLWPF4AYTN.json","view_paper":"https://pith.science/paper/WDEIK2RS","resolve_alias":"https://pith.science/api/pith-number/resolve?arxiv=2406.14833&json=true","fetch_graph":"https://pith.science/api/pith-number/WDEIK2RSZEFHKXYVQLWPF4AYTN/graph.json","fetch_events":"https://pith.science/api/pith-number/WDEIK2RSZEFHKXYVQLWPF4AYTN/events.json","actions":{"anchor_timestamp":"https://pith.science/pith/WDEIK2RSZEFHKXYVQLWPF4AYTN/action/timestamp_anchor","attest_storage":"https://pith.science/pith/WDEIK2RSZEFHKXYVQLWPF4AYTN/action/storage_attestation","attest_author":"https://pith.science/pith/WDEIK2RSZEFHKXYVQLWPF4AYTN/action/author_attestation","sign_citation":"https://pith.science/pith/WDEIK2RSZEFHKXYVQLWPF4AYTN/action/citation_signature","submit_replication":"https://pith.science/pith/WDEIK2RSZEFHKXYVQLWPF4AYTN/action/replication_record"}},"created_at":"2026-07-05T08:37:20.936692+00:00","updated_at":"2026-07-05T08:37:20.936692+00:00"}