{"record_type":"pith_number_record","schema_url":"https://pith.science/schemas/pith-number/v1.json","pith_number":"pith:2022:3TLFM7ULSZM5ZUAYKK4F46IYAD","short_pith_number":"pith:3TLFM7UL","schema_version":"1.0","canonical_sha256":"dcd6567e8b9659dcd01852b85e791800cd336d6739f6ad2775e1e19462f3db9f","source":{"kind":"arxiv","id":"2211.00635","version":3},"attestation_state":"computed","paper":{"title":"Two-stage LLM Fine-tuning with Less Specialization and More Generalization","license":"http://creativecommons.org/licenses/by/4.0/","headline":"","cross_cats":["cs.LG"],"primary_cat":"cs.CL","authors_text":"Cho-Jui Hsieh, Daliang Li, Felix Yu, Inderjit S Dhillon, Michal Lukasik, Sanjiv Kumar, Si Si, Yihan Wang","submitted_at":"2022-11-01T17:56:57Z","abstract_excerpt":"Pretrained large language models (LLMs) are general purpose problem solvers applicable to a diverse set of tasks with prompts. They can be further improved towards a specific task by fine-tuning on a specialized dataset. However, fine-tuning usually makes the model narrowly specialized on this dataset with reduced general in-context learning performances, which is undesirable whenever the fine-tuned model needs to handle additional tasks where no fine-tuning data is available. In this work, we first demonstrate that fine-tuning on a single task indeed decreases LLMs' general in-context learnin"},"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":"2211.00635","kind":"arxiv","version":3},"metadata":{"license":"http://creativecommons.org/licenses/by/4.0/","primary_cat":"cs.CL","submitted_at":"2022-11-01T17:56:57Z","cross_cats_sorted":["cs.LG"],"title_canon_sha256":"f9013531a5bff8ef155801ce943ee6889d97d62078f83a55e3f3db2d676ab965","abstract_canon_sha256":"d24612f3f19ba0e925a669dc5ac51e2f8a7eac9a7ea8e714047e4ec8018de029"},"schema_version":"1.0"},"receipt":{"kind":"pith_receipt","key_id":"pith-v1-2026-05","algorithm":"ed25519","signed_at":"2026-07-05T07:55:19.227419Z","signature_b64":"lk2wG1Z9HekmJ5qMvJZ0vBPMzW0kiGWX34ycRldExixK97uWESaj8fTInolwl0ujBORt8WlAUYtS8br0b0cmBA==","signed_message":"canonical_sha256_bytes","builder_version":"pith-number-builder-2026-05-17-v1","receipt_version":"0.3","canonical_sha256":"dcd6567e8b9659dcd01852b85e791800cd336d6739f6ad2775e1e19462f3db9f","last_reissued_at":"2026-07-05T07:55:19.226889Z","signature_status":"signed_v1","first_computed_at":"2026-07-05T07:55:19.226889Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"graph_snapshot":{"paper":{"title":"Two-stage LLM Fine-tuning with Less Specialization and More Generalization","license":"http://creativecommons.org/licenses/by/4.0/","headline":"","cross_cats":["cs.LG"],"primary_cat":"cs.CL","authors_text":"Cho-Jui Hsieh, Daliang Li, Felix Yu, Inderjit S Dhillon, Michal Lukasik, Sanjiv Kumar, Si Si, Yihan Wang","submitted_at":"2022-11-01T17:56:57Z","abstract_excerpt":"Pretrained large language models (LLMs) are general purpose problem solvers applicable to a diverse set of tasks with prompts. They can be further improved towards a specific task by fine-tuning on a specialized dataset. However, fine-tuning usually makes the model narrowly specialized on this dataset with reduced general in-context learning performances, which is undesirable whenever the fine-tuned model needs to handle additional tasks where no fine-tuning data is available. In this work, we first demonstrate that fine-tuning on a single task indeed decreases LLMs' general in-context learnin"},"claims":{"count":0,"items":[],"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"source":{"id":"2211.00635","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/2211.00635/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":"2211.00635","created_at":"2026-07-05T07:55:19.226948+00:00"},{"alias_kind":"arxiv_version","alias_value":"2211.00635v3","created_at":"2026-07-05T07:55:19.226948+00:00"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2211.00635","created_at":"2026-07-05T07:55:19.226948+00:00"},{"alias_kind":"pith_short_12","alias_value":"3TLFM7ULSZM5","created_at":"2026-07-05T07:55:19.226948+00:00"},{"alias_kind":"pith_short_16","alias_value":"3TLFM7ULSZM5ZUAY","created_at":"2026-07-05T07:55:19.226948+00:00"},{"alias_kind":"pith_short_8","alias_value":"3TLFM7UL","created_at":"2026-07-05T07:55:19.226948+00:00"}],"events":[],"event_summary":{},"paper_claims":[],"inbound_citations":{"count":2,"internal_anchor_count":0,"sample":[{"citing_arxiv_id":"2605.17967","citing_title":"Reconciling Contradictory Views on the Effectiveness of SFT in LLMs: An Interaction Perspective","ref_index":35,"is_internal_anchor":false},{"citing_arxiv_id":"2605.11922","citing_title":"StepCodeReasoner: Aligning Code Reasoning with Stepwise Execution Traces via Reinforcement Learning","ref_index":64,"is_internal_anchor":false}]},"formal_canon":{"evidence_count":0,"sample":[],"anchors":[]},"links":{"html":"https://pith.science/pith/3TLFM7ULSZM5ZUAYKK4F46IYAD","json":"https://pith.science/pith/3TLFM7ULSZM5ZUAYKK4F46IYAD.json","graph_json":"https://pith.science/api/pith-number/3TLFM7ULSZM5ZUAYKK4F46IYAD/graph.json","events_json":"https://pith.science/api/pith-number/3TLFM7ULSZM5ZUAYKK4F46IYAD/events.json","paper":"https://pith.science/paper/3TLFM7UL"},"agent_actions":{"view_html":"https://pith.science/pith/3TLFM7ULSZM5ZUAYKK4F46IYAD","download_json":"https://pith.science/pith/3TLFM7ULSZM5ZUAYKK4F46IYAD.json","view_paper":"https://pith.science/paper/3TLFM7UL","resolve_alias":"https://pith.science/api/pith-number/resolve?arxiv=2211.00635&json=true","fetch_graph":"https://pith.science/api/pith-number/3TLFM7ULSZM5ZUAYKK4F46IYAD/graph.json","fetch_events":"https://pith.science/api/pith-number/3TLFM7ULSZM5ZUAYKK4F46IYAD/events.json","actions":{"anchor_timestamp":"https://pith.science/pith/3TLFM7ULSZM5ZUAYKK4F46IYAD/action/timestamp_anchor","attest_storage":"https://pith.science/pith/3TLFM7ULSZM5ZUAYKK4F46IYAD/action/storage_attestation","attest_author":"https://pith.science/pith/3TLFM7ULSZM5ZUAYKK4F46IYAD/action/author_attestation","sign_citation":"https://pith.science/pith/3TLFM7ULSZM5ZUAYKK4F46IYAD/action/citation_signature","submit_replication":"https://pith.science/pith/3TLFM7ULSZM5ZUAYKK4F46IYAD/action/replication_record"}},"created_at":"2026-07-05T07:55:19.226948+00:00","updated_at":"2026-07-05T07:55:19.226948+00:00"}