{"record_type":"pith_number_record","schema_url":"https://pith.science/schemas/pith-number/v1.json","pith_number":"pith:2024:BXY6Z5YIIIUPOIUYI24OEYINHR","short_pith_number":"pith:BXY6Z5YI","schema_version":"1.0","canonical_sha256":"0df1ecf7084228f7229846b8e2610d3c47610f9b152ee5756b5145ef0ef0fadd","source":{"kind":"arxiv","id":"2410.07471","version":2},"attestation_state":"computed","paper":{"title":"SEAL: Safety-enhanced Aligned LLM Fine-tuning via Bilevel Data Selection","license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","headline":"","cross_cats":["cs.AI","cs.CL"],"primary_cat":"cs.LG","authors_text":"Han Shen, Payel Das, Pin-Yu Chen, Tianyi Chen","submitted_at":"2024-10-09T22:24:22Z","abstract_excerpt":"Fine-tuning on task-specific data to boost downstream performance is a crucial step for leveraging Large Language Models (LLMs). However, previous studies have demonstrated that fine-tuning the models on several adversarial samples or even benign data can greatly comprise the model's pre-equipped alignment and safety capabilities. In this work, we propose SEAL, a novel framework to enhance safety in LLM fine-tuning. SEAL learns a data ranker based on the bilevel optimization to up rank the safe and high-quality fine-tuning data and down rank the unsafe or low-quality ones. Models trained with "},"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":"2410.07471","kind":"arxiv","version":2},"metadata":{"license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","primary_cat":"cs.LG","submitted_at":"2024-10-09T22:24:22Z","cross_cats_sorted":["cs.AI","cs.CL"],"title_canon_sha256":"2fcabf01881c1e39ab9b9a722376a470b0732eefd592682684fecc72fba29221","abstract_canon_sha256":"b51dceff7c7b916fa158865bcabf3830a8aa7a2438c1dfac665d497fb43f86d0"},"schema_version":"1.0"},"receipt":{"kind":"pith_receipt","key_id":"pith-v1-2026-05","algorithm":"ed25519","signed_at":"2026-07-05T09:18:54.465274Z","signature_b64":"xwwembS/3Ytqb8WPFyggroAgzBLRmT6OnT7EK6j6BjKBHE7dzp6RpccKLKvSu29THjeqGXA2LAfD5o8xL36jDw==","signed_message":"canonical_sha256_bytes","builder_version":"pith-number-builder-2026-05-17-v1","receipt_version":"0.3","canonical_sha256":"0df1ecf7084228f7229846b8e2610d3c47610f9b152ee5756b5145ef0ef0fadd","last_reissued_at":"2026-07-05T09:18:54.464794Z","signature_status":"signed_v1","first_computed_at":"2026-07-05T09:18:54.464794Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"graph_snapshot":{"paper":{"title":"SEAL: Safety-enhanced Aligned LLM Fine-tuning via Bilevel Data Selection","license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","headline":"","cross_cats":["cs.AI","cs.CL"],"primary_cat":"cs.LG","authors_text":"Han Shen, Payel Das, Pin-Yu Chen, Tianyi Chen","submitted_at":"2024-10-09T22:24:22Z","abstract_excerpt":"Fine-tuning on task-specific data to boost downstream performance is a crucial step for leveraging Large Language Models (LLMs). However, previous studies have demonstrated that fine-tuning the models on several adversarial samples or even benign data can greatly comprise the model's pre-equipped alignment and safety capabilities. In this work, we propose SEAL, a novel framework to enhance safety in LLM fine-tuning. SEAL learns a data ranker based on the bilevel optimization to up rank the safe and high-quality fine-tuning data and down rank the unsafe or low-quality ones. Models trained with "},"claims":{"count":0,"items":[],"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"source":{"id":"2410.07471","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/2410.07471/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":"2410.07471","created_at":"2026-07-05T09:18:54.464853+00:00"},{"alias_kind":"arxiv_version","alias_value":"2410.07471v2","created_at":"2026-07-05T09:18:54.464853+00:00"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2410.07471","created_at":"2026-07-05T09:18:54.464853+00:00"},{"alias_kind":"pith_short_12","alias_value":"BXY6Z5YIIIUP","created_at":"2026-07-05T09:18:54.464853+00:00"},{"alias_kind":"pith_short_16","alias_value":"BXY6Z5YIIIUPOIUY","created_at":"2026-07-05T09:18:54.464853+00:00"},{"alias_kind":"pith_short_8","alias_value":"BXY6Z5YI","created_at":"2026-07-05T09:18:54.464853+00:00"}],"events":[],"event_summary":{},"paper_claims":[],"inbound_citations":{"count":5,"internal_anchor_count":0,"sample":[{"citing_arxiv_id":"2606.30263","citing_title":"Defending Against Harmful Supervision Hidden in Benign Samples","ref_index":5,"is_internal_anchor":false},{"citing_arxiv_id":"2605.28664","citing_title":"Activation Steering for Synthetic Data Generation: The Role of Diversity in Downstream Safety Detection","ref_index":45,"is_internal_anchor":false},{"citing_arxiv_id":"2409.18169","citing_title":"Harmful Fine-tuning Attacks and Defenses for Large Language Models: A Survey","ref_index":133,"is_internal_anchor":false},{"citing_arxiv_id":"2605.10288","citing_title":"BROS: Bias-Corrected Randomized Subspaces for Memory-Efficient Single-Loop Bilevel Optimization","ref_index":50,"is_internal_anchor":false},{"citing_arxiv_id":"2605.10288","citing_title":"BROS: Bias-Corrected Randomized Subspaces for Memory-Efficient Single-Loop Bilevel Optimization","ref_index":50,"is_internal_anchor":false}]},"formal_canon":{"evidence_count":0,"sample":[],"anchors":[]},"links":{"html":"https://pith.science/pith/BXY6Z5YIIIUPOIUYI24OEYINHR","json":"https://pith.science/pith/BXY6Z5YIIIUPOIUYI24OEYINHR.json","graph_json":"https://pith.science/api/pith-number/BXY6Z5YIIIUPOIUYI24OEYINHR/graph.json","events_json":"https://pith.science/api/pith-number/BXY6Z5YIIIUPOIUYI24OEYINHR/events.json","paper":"https://pith.science/paper/BXY6Z5YI"},"agent_actions":{"view_html":"https://pith.science/pith/BXY6Z5YIIIUPOIUYI24OEYINHR","download_json":"https://pith.science/pith/BXY6Z5YIIIUPOIUYI24OEYINHR.json","view_paper":"https://pith.science/paper/BXY6Z5YI","resolve_alias":"https://pith.science/api/pith-number/resolve?arxiv=2410.07471&json=true","fetch_graph":"https://pith.science/api/pith-number/BXY6Z5YIIIUPOIUYI24OEYINHR/graph.json","fetch_events":"https://pith.science/api/pith-number/BXY6Z5YIIIUPOIUYI24OEYINHR/events.json","actions":{"anchor_timestamp":"https://pith.science/pith/BXY6Z5YIIIUPOIUYI24OEYINHR/action/timestamp_anchor","attest_storage":"https://pith.science/pith/BXY6Z5YIIIUPOIUYI24OEYINHR/action/storage_attestation","attest_author":"https://pith.science/pith/BXY6Z5YIIIUPOIUYI24OEYINHR/action/author_attestation","sign_citation":"https://pith.science/pith/BXY6Z5YIIIUPOIUYI24OEYINHR/action/citation_signature","submit_replication":"https://pith.science/pith/BXY6Z5YIIIUPOIUYI24OEYINHR/action/replication_record"}},"created_at":"2026-07-05T09:18:54.464853+00:00","updated_at":"2026-07-05T09:18:54.464853+00:00"}