{"record_type":"pith_number_record","schema_url":"https://pith.science/schemas/pith-number/v1.json","pith_number":"pith:2025:IGC4CW72CYGSXQIVUWRXCQYF6E","short_pith_number":"pith:IGC4CW72","schema_version":"1.0","canonical_sha256":"4185c15bfa160d2bc115a5a3714305f10147f4c8fcfe5e59db680fe8362e8edb","source":{"kind":"arxiv","id":"2505.12038","version":1},"attestation_state":"computed","paper":{"title":"Safe Delta: Consistently Preserving Safety when Fine-Tuning LLMs on Diverse Datasets","license":"http://creativecommons.org/licenses/by-nc-nd/4.0/","headline":"","cross_cats":["cs.AI","cs.CR"],"primary_cat":"cs.LG","authors_text":"Jiahao Wu, Ke Tang, Ning Lu, Qi Wang, Shengcai Liu, Weiyu Chen, Yew-Soon Ong, Zhirui Zhang","submitted_at":"2025-05-17T15:01:07Z","abstract_excerpt":"Large language models (LLMs) have shown great potential as general-purpose AI assistants across various domains. To fully leverage this potential in specific applications, many companies provide fine-tuning API services, enabling users to upload their own data for LLM customization. However, fine-tuning services introduce a new safety threat: user-uploaded data, whether harmful or benign, can break the model's alignment, leading to unsafe outputs. Moreover, existing defense methods struggle to address the diversity of fine-tuning datasets (e.g., varying sizes, tasks), often sacrificing utility"},"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":"2505.12038","kind":"arxiv","version":1},"metadata":{"license":"http://creativecommons.org/licenses/by-nc-nd/4.0/","primary_cat":"cs.LG","submitted_at":"2025-05-17T15:01:07Z","cross_cats_sorted":["cs.AI","cs.CR"],"title_canon_sha256":"6b8354b7810e525094b1b47fa1fd438bb38b433eecc077fed439a0b2f0625f55","abstract_canon_sha256":"3b83f4702c2a06a0eaa2bd903b13ffee75c458da328649601430199b54dc5659"},"schema_version":"1.0"},"receipt":{"kind":"pith_receipt","key_id":"pith-v1-2026-05","algorithm":"ed25519","signed_at":"2026-07-05T11:04:57.374283Z","signature_b64":"Bq9thu+SELWqkKEPAVedDSMVOxOKglHF7jEXUafXYjrimH6g3DkGazLVHjSCwoE2Ch3OCtvjbRb5iZGk7NrIBQ==","signed_message":"canonical_sha256_bytes","builder_version":"pith-number-builder-2026-05-17-v1","receipt_version":"0.3","canonical_sha256":"4185c15bfa160d2bc115a5a3714305f10147f4c8fcfe5e59db680fe8362e8edb","last_reissued_at":"2026-07-05T11:04:57.373755Z","signature_status":"signed_v1","first_computed_at":"2026-07-05T11:04:57.373755Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"graph_snapshot":{"paper":{"title":"Safe Delta: Consistently Preserving Safety when Fine-Tuning LLMs on Diverse Datasets","license":"http://creativecommons.org/licenses/by-nc-nd/4.0/","headline":"","cross_cats":["cs.AI","cs.CR"],"primary_cat":"cs.LG","authors_text":"Jiahao Wu, Ke Tang, Ning Lu, Qi Wang, Shengcai Liu, Weiyu Chen, Yew-Soon Ong, Zhirui Zhang","submitted_at":"2025-05-17T15:01:07Z","abstract_excerpt":"Large language models (LLMs) have shown great potential as general-purpose AI assistants across various domains. To fully leverage this potential in specific applications, many companies provide fine-tuning API services, enabling users to upload their own data for LLM customization. However, fine-tuning services introduce a new safety threat: user-uploaded data, whether harmful or benign, can break the model's alignment, leading to unsafe outputs. Moreover, existing defense methods struggle to address the diversity of fine-tuning datasets (e.g., varying sizes, tasks), often sacrificing utility"},"claims":{"count":0,"items":[],"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"source":{"id":"2505.12038","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/2505.12038/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":"2505.12038","created_at":"2026-07-05T11:04:57.373830+00:00"},{"alias_kind":"arxiv_version","alias_value":"2505.12038v1","created_at":"2026-07-05T11:04:57.373830+00:00"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2505.12038","created_at":"2026-07-05T11:04:57.373830+00:00"},{"alias_kind":"pith_short_12","alias_value":"IGC4CW72CYGS","created_at":"2026-07-05T11:04:57.373830+00:00"},{"alias_kind":"pith_short_16","alias_value":"IGC4CW72CYGSXQIV","created_at":"2026-07-05T11:04:57.373830+00:00"},{"alias_kind":"pith_short_8","alias_value":"IGC4CW72","created_at":"2026-07-05T11:04:57.373830+00:00"}],"events":[],"event_summary":{},"paper_claims":[],"inbound_citations":{"count":1,"internal_anchor_count":0,"sample":[{"citing_arxiv_id":"2604.17215","citing_title":"Continual Safety Alignment via Gradient-Based Sample Selection","ref_index":2,"is_internal_anchor":false}]},"formal_canon":{"evidence_count":0,"sample":[],"anchors":[]},"links":{"html":"https://pith.science/pith/IGC4CW72CYGSXQIVUWRXCQYF6E","json":"https://pith.science/pith/IGC4CW72CYGSXQIVUWRXCQYF6E.json","graph_json":"https://pith.science/api/pith-number/IGC4CW72CYGSXQIVUWRXCQYF6E/graph.json","events_json":"https://pith.science/api/pith-number/IGC4CW72CYGSXQIVUWRXCQYF6E/events.json","paper":"https://pith.science/paper/IGC4CW72"},"agent_actions":{"view_html":"https://pith.science/pith/IGC4CW72CYGSXQIVUWRXCQYF6E","download_json":"https://pith.science/pith/IGC4CW72CYGSXQIVUWRXCQYF6E.json","view_paper":"https://pith.science/paper/IGC4CW72","resolve_alias":"https://pith.science/api/pith-number/resolve?arxiv=2505.12038&json=true","fetch_graph":"https://pith.science/api/pith-number/IGC4CW72CYGSXQIVUWRXCQYF6E/graph.json","fetch_events":"https://pith.science/api/pith-number/IGC4CW72CYGSXQIVUWRXCQYF6E/events.json","actions":{"anchor_timestamp":"https://pith.science/pith/IGC4CW72CYGSXQIVUWRXCQYF6E/action/timestamp_anchor","attest_storage":"https://pith.science/pith/IGC4CW72CYGSXQIVUWRXCQYF6E/action/storage_attestation","attest_author":"https://pith.science/pith/IGC4CW72CYGSXQIVUWRXCQYF6E/action/author_attestation","sign_citation":"https://pith.science/pith/IGC4CW72CYGSXQIVUWRXCQYF6E/action/citation_signature","submit_replication":"https://pith.science/pith/IGC4CW72CYGSXQIVUWRXCQYF6E/action/replication_record"}},"created_at":"2026-07-05T11:04:57.373830+00:00","updated_at":"2026-07-05T11:04:57.373830+00:00"}