{"record_type":"pith_number_record","schema_url":"https://pith.science/schemas/pith-number/v1.json","pith_number":"pith:2024:3ZKQA6NHBVEJU6UEG2KPBLI7CG","short_pith_number":"pith:3ZKQA6NH","schema_version":"1.0","canonical_sha256":"de550079a70d489a7a843694f0ad1f1188d492a9192e2dd38301720edf6fd6e1","source":{"kind":"arxiv","id":"2402.10193","version":3},"attestation_state":"computed","paper":{"title":"BitDelta: Your Fine-Tune May Only Be Worth One Bit","license":"http://creativecommons.org/licenses/by/4.0/","headline":"","cross_cats":["cs.CL"],"primary_cat":"cs.LG","authors_text":"Guangxuan Xiao, James Liu, Jason D. Lee, Kai Li, Song Han, Tianle Cai, Tri Dao","submitted_at":"2024-02-15T18:50:06Z","abstract_excerpt":"Large Language Models (LLMs) are typically trained in two phases: pre-training on large internet-scale datasets, and fine-tuning for downstream tasks. Given the higher computational demand of pre-training, it's intuitive to assume that fine-tuning adds less new information to the model, and is thus more compressible. We explore this assumption by decomposing the weights of fine-tuned models into their pre-trained components and an additional delta. We introduce a simple method, BitDelta, which successfully quantizes this delta down to 1 bit without compromising performance. This interesting fi"},"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":"2402.10193","kind":"arxiv","version":3},"metadata":{"license":"http://creativecommons.org/licenses/by/4.0/","primary_cat":"cs.LG","submitted_at":"2024-02-15T18:50:06Z","cross_cats_sorted":["cs.CL"],"title_canon_sha256":"631409bf256bcf3a8e4329990c9ee58a1ac94b46f78c3f5a635e0b24c47211d8","abstract_canon_sha256":"deb3294f02fded9be61aac9b51393a1b5cc4b548b7f87f1f6827d3c0e31b4a75"},"schema_version":"1.0"},"receipt":{"kind":"pith_receipt","key_id":"pith-v1-2026-05","algorithm":"ed25519","signed_at":"2026-07-05T09:19:54.185500Z","signature_b64":"501oslYxelBYGeWSL49RwYCUvPfzYCWStwqjR9RYxz9AgWYpJbcexyQ/Thriv7Y/V4FBY7r1x+fXITpwRm0rDw==","signed_message":"canonical_sha256_bytes","builder_version":"pith-number-builder-2026-05-17-v1","receipt_version":"0.3","canonical_sha256":"de550079a70d489a7a843694f0ad1f1188d492a9192e2dd38301720edf6fd6e1","last_reissued_at":"2026-07-05T09:19:54.185021Z","signature_status":"signed_v1","first_computed_at":"2026-07-05T09:19:54.185021Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"graph_snapshot":{"paper":{"title":"BitDelta: Your Fine-Tune May Only Be Worth One Bit","license":"http://creativecommons.org/licenses/by/4.0/","headline":"","cross_cats":["cs.CL"],"primary_cat":"cs.LG","authors_text":"Guangxuan Xiao, James Liu, Jason D. Lee, Kai Li, Song Han, Tianle Cai, Tri Dao","submitted_at":"2024-02-15T18:50:06Z","abstract_excerpt":"Large Language Models (LLMs) are typically trained in two phases: pre-training on large internet-scale datasets, and fine-tuning for downstream tasks. Given the higher computational demand of pre-training, it's intuitive to assume that fine-tuning adds less new information to the model, and is thus more compressible. We explore this assumption by decomposing the weights of fine-tuned models into their pre-trained components and an additional delta. We introduce a simple method, BitDelta, which successfully quantizes this delta down to 1 bit without compromising performance. This interesting fi"},"claims":{"count":0,"items":[],"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"source":{"id":"2402.10193","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/2402.10193/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":"2402.10193","created_at":"2026-07-05T09:19:54.185077+00:00"},{"alias_kind":"arxiv_version","alias_value":"2402.10193v3","created_at":"2026-07-05T09:19:54.185077+00:00"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2402.10193","created_at":"2026-07-05T09:19:54.185077+00:00"},{"alias_kind":"pith_short_12","alias_value":"3ZKQA6NHBVEJ","created_at":"2026-07-05T09:19:54.185077+00:00"},{"alias_kind":"pith_short_16","alias_value":"3ZKQA6NHBVEJU6UE","created_at":"2026-07-05T09:19:54.185077+00:00"},{"alias_kind":"pith_short_8","alias_value":"3ZKQA6NH","created_at":"2026-07-05T09:19:54.185077+00:00"}],"events":[],"event_summary":{},"paper_claims":[],"inbound_citations":{"count":2,"internal_anchor_count":0,"sample":[{"citing_arxiv_id":"2403.14608","citing_title":"Parameter-Efficient Fine-Tuning for Large Models: A Comprehensive Survey","ref_index":130,"is_internal_anchor":false},{"citing_arxiv_id":"2604.24273","citing_title":"BitRL: Reinforcement Learning with 1-bit Quantized Language Models for Resource-Constrained Edge Deployment","ref_index":6,"is_internal_anchor":false}]},"formal_canon":{"evidence_count":0,"sample":[],"anchors":[]},"links":{"html":"https://pith.science/pith/3ZKQA6NHBVEJU6UEG2KPBLI7CG","json":"https://pith.science/pith/3ZKQA6NHBVEJU6UEG2KPBLI7CG.json","graph_json":"https://pith.science/api/pith-number/3ZKQA6NHBVEJU6UEG2KPBLI7CG/graph.json","events_json":"https://pith.science/api/pith-number/3ZKQA6NHBVEJU6UEG2KPBLI7CG/events.json","paper":"https://pith.science/paper/3ZKQA6NH"},"agent_actions":{"view_html":"https://pith.science/pith/3ZKQA6NHBVEJU6UEG2KPBLI7CG","download_json":"https://pith.science/pith/3ZKQA6NHBVEJU6UEG2KPBLI7CG.json","view_paper":"https://pith.science/paper/3ZKQA6NH","resolve_alias":"https://pith.science/api/pith-number/resolve?arxiv=2402.10193&json=true","fetch_graph":"https://pith.science/api/pith-number/3ZKQA6NHBVEJU6UEG2KPBLI7CG/graph.json","fetch_events":"https://pith.science/api/pith-number/3ZKQA6NHBVEJU6UEG2KPBLI7CG/events.json","actions":{"anchor_timestamp":"https://pith.science/pith/3ZKQA6NHBVEJU6UEG2KPBLI7CG/action/timestamp_anchor","attest_storage":"https://pith.science/pith/3ZKQA6NHBVEJU6UEG2KPBLI7CG/action/storage_attestation","attest_author":"https://pith.science/pith/3ZKQA6NHBVEJU6UEG2KPBLI7CG/action/author_attestation","sign_citation":"https://pith.science/pith/3ZKQA6NHBVEJU6UEG2KPBLI7CG/action/citation_signature","submit_replication":"https://pith.science/pith/3ZKQA6NHBVEJU6UEG2KPBLI7CG/action/replication_record"}},"created_at":"2026-07-05T09:19:54.185077+00:00","updated_at":"2026-07-05T09:19:54.185077+00:00"}