{"record_type":"pith_number_record","schema_url":"https://pith.science/schemas/pith-number/v1.json","pith_number":"pith:2024:T2WLR2RBRBJYWHO55JDR5E55N2","short_pith_number":"pith:T2WLR2RB","schema_version":"1.0","canonical_sha256":"9eacb8ea2188538b1dddea471e93bd6e83de4768aa22d071733f71f7b47c66ae","source":{"kind":"arxiv","id":"2406.14785","version":1},"attestation_state":"computed","paper":{"title":"Understanding Finetuning for Factual Knowledge Extraction","license":"http://creativecommons.org/licenses/by/4.0/","headline":"","cross_cats":["cs.LG"],"primary_cat":"cs.CL","authors_text":"Aditi Raghunathan, Gaurav Ghosal, Tatsunori Hashimoto","submitted_at":"2024-06-20T23:27:06Z","abstract_excerpt":"In this work, we study the impact of QA fine-tuning data on downstream factuality. We show that fine-tuning on lesser-known facts that are poorly stored during pretraining yields significantly worse factuality than fine-tuning on well-known facts, even when all facts are seen during pretraining. We prove this phenomenon theoretically, showing that training on lesser-known facts can lead the model to ignore subject entity names and instead output a generic plausible response even when the relevant factual knowledge is encoded in the model. On three question answering benchmarks (PopQA, Entity Q"},"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":"2406.14785","kind":"arxiv","version":1},"metadata":{"license":"http://creativecommons.org/licenses/by/4.0/","primary_cat":"cs.CL","submitted_at":"2024-06-20T23:27:06Z","cross_cats_sorted":["cs.LG"],"title_canon_sha256":"ec624b198c39e2dfb46f4fb898162bc1fed97ab40bc75f79c299ab244cd891f5","abstract_canon_sha256":"8c2ef70863692aae3b9dd1ff1c9e61985b9361146c24a04cbc2f2b7d84a04a4e"},"schema_version":"1.0"},"receipt":{"kind":"pith_receipt","key_id":"pith-v1-2026-05","algorithm":"ed25519","signed_at":"2026-07-05T08:35:08.383475Z","signature_b64":"hmglRJcY8kjF97d7D3iCC7WTpa1ZUexnnlvaAjYpii7c5Zojj+/JIq2i318SROBUlkZeuCAA/4yjRQP6Ozb7DQ==","signed_message":"canonical_sha256_bytes","builder_version":"pith-number-builder-2026-05-17-v1","receipt_version":"0.3","canonical_sha256":"9eacb8ea2188538b1dddea471e93bd6e83de4768aa22d071733f71f7b47c66ae","last_reissued_at":"2026-07-05T08:35:08.382934Z","signature_status":"signed_v1","first_computed_at":"2026-07-05T08:35:08.382934Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"graph_snapshot":{"paper":{"title":"Understanding Finetuning for Factual Knowledge Extraction","license":"http://creativecommons.org/licenses/by/4.0/","headline":"","cross_cats":["cs.LG"],"primary_cat":"cs.CL","authors_text":"Aditi Raghunathan, Gaurav Ghosal, Tatsunori Hashimoto","submitted_at":"2024-06-20T23:27:06Z","abstract_excerpt":"In this work, we study the impact of QA fine-tuning data on downstream factuality. We show that fine-tuning on lesser-known facts that are poorly stored during pretraining yields significantly worse factuality than fine-tuning on well-known facts, even when all facts are seen during pretraining. We prove this phenomenon theoretically, showing that training on lesser-known facts can lead the model to ignore subject entity names and instead output a generic plausible response even when the relevant factual knowledge is encoded in the model. On three question answering benchmarks (PopQA, Entity Q"},"claims":{"count":0,"items":[],"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"source":{"id":"2406.14785","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/2406.14785/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":"2406.14785","created_at":"2026-07-05T08:35:08.382990+00:00"},{"alias_kind":"arxiv_version","alias_value":"2406.14785v1","created_at":"2026-07-05T08:35:08.382990+00:00"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2406.14785","created_at":"2026-07-05T08:35:08.382990+00:00"},{"alias_kind":"pith_short_12","alias_value":"T2WLR2RBRBJY","created_at":"2026-07-05T08:35:08.382990+00:00"},{"alias_kind":"pith_short_16","alias_value":"T2WLR2RBRBJYWHO5","created_at":"2026-07-05T08:35:08.382990+00:00"},{"alias_kind":"pith_short_8","alias_value":"T2WLR2RB","created_at":"2026-07-05T08:35:08.382990+00:00"}],"events":[],"event_summary":{},"paper_claims":[],"inbound_citations":{"count":2,"internal_anchor_count":0,"sample":[{"citing_arxiv_id":"2508.00901","citing_title":"Provable Knowledge Acquisition and Extraction in One-Layer Transformers","ref_index":12,"is_internal_anchor":false},{"citing_arxiv_id":"2511.02626","citing_title":"Understanding New-Knowledge-Induced Factual Hallucinations in LLMs: Analysis and Interpretation","ref_index":8,"is_internal_anchor":false}]},"formal_canon":{"evidence_count":0,"sample":[],"anchors":[]},"links":{"html":"https://pith.science/pith/T2WLR2RBRBJYWHO55JDR5E55N2","json":"https://pith.science/pith/T2WLR2RBRBJYWHO55JDR5E55N2.json","graph_json":"https://pith.science/api/pith-number/T2WLR2RBRBJYWHO55JDR5E55N2/graph.json","events_json":"https://pith.science/api/pith-number/T2WLR2RBRBJYWHO55JDR5E55N2/events.json","paper":"https://pith.science/paper/T2WLR2RB"},"agent_actions":{"view_html":"https://pith.science/pith/T2WLR2RBRBJYWHO55JDR5E55N2","download_json":"https://pith.science/pith/T2WLR2RBRBJYWHO55JDR5E55N2.json","view_paper":"https://pith.science/paper/T2WLR2RB","resolve_alias":"https://pith.science/api/pith-number/resolve?arxiv=2406.14785&json=true","fetch_graph":"https://pith.science/api/pith-number/T2WLR2RBRBJYWHO55JDR5E55N2/graph.json","fetch_events":"https://pith.science/api/pith-number/T2WLR2RBRBJYWHO55JDR5E55N2/events.json","actions":{"anchor_timestamp":"https://pith.science/pith/T2WLR2RBRBJYWHO55JDR5E55N2/action/timestamp_anchor","attest_storage":"https://pith.science/pith/T2WLR2RBRBJYWHO55JDR5E55N2/action/storage_attestation","attest_author":"https://pith.science/pith/T2WLR2RBRBJYWHO55JDR5E55N2/action/author_attestation","sign_citation":"https://pith.science/pith/T2WLR2RBRBJYWHO55JDR5E55N2/action/citation_signature","submit_replication":"https://pith.science/pith/T2WLR2RBRBJYWHO55JDR5E55N2/action/replication_record"}},"created_at":"2026-07-05T08:35:08.382990+00:00","updated_at":"2026-07-05T08:35:08.382990+00:00"}