{"record_type":"pith_number_record","schema_url":"https://pith.science/schemas/pith-number/v1.json","pith_number":"pith:2024:DIFWTZLUGIWN5BBP5AMRODX7Y6","short_pith_number":"pith:DIFWTZLU","schema_version":"1.0","canonical_sha256":"1a0b69e574322cde842fe819170effc79657bbd7828125f3c72865f4b9972bee","source":{"kind":"arxiv","id":"2406.15513","version":3},"attestation_state":"computed","paper":{"title":"PKU-SafeRLHF: Towards Multi-Level Safety Alignment for LLMs with Human Preference","license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","headline":"","cross_cats":["cs.CL"],"primary_cat":"cs.AI","authors_text":"Boren Zheng, Borong Zhang, Boxuan Li, Boyuan Chen, Donghai Hong, Jiaming Ji, Jiayi Zhou, Juntao Dai, Kaile Wang, Sirui Han, Tianyi Qiu, Yaodong Yang, Yike Guo","submitted_at":"2024-06-20T18:37:36Z","abstract_excerpt":"In this study, we introduce the safety human preference dataset, PKU-SafeRLHF, designed to promote research on safety alignment in large language models (LLMs). As a sibling project to SafeRLHF and BeaverTails, we separate annotations of helpfulness and harmlessness for question-answering pairs, providing distinct perspectives on these coupled attributes. Overall, we provide 44.6k refined prompts and 265k question-answer pairs with safety meta-labels for 19 harm categories and three severity levels ranging from minor to severe, with answers generated by Llama-family models. Based on this, we c"},"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":"2406.15513","kind":"arxiv","version":3},"metadata":{"license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","primary_cat":"cs.AI","submitted_at":"2024-06-20T18:37:36Z","cross_cats_sorted":["cs.CL"],"title_canon_sha256":"c47e417c767a0ef1d2be356e4446dc852cec3a9f68110f8e142366affd34d4c1","abstract_canon_sha256":"7c96a6d6ef28ec700958df9192a7ce2e495806840e3f10fa4450956814cee8bf"},"schema_version":"1.0"},"receipt":{"kind":"pith_receipt","key_id":"pith-v1-2026-05","algorithm":"ed25519","signed_at":"2026-07-05T11:21:33.639484Z","signature_b64":"LU+RuOnK8w8ArZ6pstIsWoGicrWBInZ++XQdQ2utfwVuv4Zgk0im6KulURWckBkaZYSGfMW5npWvBz2SClaEBA==","signed_message":"canonical_sha256_bytes","builder_version":"pith-number-builder-2026-05-17-v1","receipt_version":"0.3","canonical_sha256":"1a0b69e574322cde842fe819170effc79657bbd7828125f3c72865f4b9972bee","last_reissued_at":"2026-07-05T11:21:33.638947Z","signature_status":"signed_v1","first_computed_at":"2026-07-05T11:21:33.638947Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"graph_snapshot":{"paper":{"title":"PKU-SafeRLHF: Towards Multi-Level Safety Alignment for LLMs with Human Preference","license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","headline":"","cross_cats":["cs.CL"],"primary_cat":"cs.AI","authors_text":"Boren Zheng, Borong Zhang, Boxuan Li, Boyuan Chen, Donghai Hong, Jiaming Ji, Jiayi Zhou, Juntao Dai, Kaile Wang, Sirui Han, Tianyi Qiu, Yaodong Yang, Yike Guo","submitted_at":"2024-06-20T18:37:36Z","abstract_excerpt":"In this study, we introduce the safety human preference dataset, PKU-SafeRLHF, designed to promote research on safety alignment in large language models (LLMs). As a sibling project to SafeRLHF and BeaverTails, we separate annotations of helpfulness and harmlessness for question-answering pairs, providing distinct perspectives on these coupled attributes. Overall, we provide 44.6k refined prompts and 265k question-answer pairs with safety meta-labels for 19 harm categories and three severity levels ranging from minor to severe, with answers generated by Llama-family models. Based on this, we c"},"claims":{"count":0,"items":[],"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"source":{"id":"2406.15513","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/2406.15513/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":"2406.15513","created_at":"2026-07-05T11:21:33.639010+00:00"},{"alias_kind":"arxiv_version","alias_value":"2406.15513v3","created_at":"2026-07-05T11:21:33.639010+00:00"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2406.15513","created_at":"2026-07-05T11:21:33.639010+00:00"},{"alias_kind":"pith_short_12","alias_value":"DIFWTZLUGIWN","created_at":"2026-07-05T11:21:33.639010+00:00"},{"alias_kind":"pith_short_16","alias_value":"DIFWTZLUGIWN5BBP","created_at":"2026-07-05T11:21:33.639010+00:00"},{"alias_kind":"pith_short_8","alias_value":"DIFWTZLU","created_at":"2026-07-05T11:21:33.639010+00:00"}],"events":[],"event_summary":{},"paper_claims":[],"inbound_citations":{"count":26,"internal_anchor_count":1,"sample":[{"citing_arxiv_id":"2607.07907","citing_title":"Multimodal Unlearning Across Vision, Language, Video, and Audio: Survey of Methods, Datasets, and Benchmarks","ref_index":276,"is_internal_anchor":true},{"citing_arxiv_id":"2607.02047","citing_title":"OpenSafeIntent: Evaluating Intent-Calibrated Safe Completion Across Dual-Use Prompt Sets","ref_index":23,"is_internal_anchor":false},{"citing_arxiv_id":"2606.08044","citing_title":"When Behavioral Safety Evaluation Fails: A Representation-Level Perspective","ref_index":17,"is_internal_anchor":false},{"citing_arxiv_id":"2606.07335","citing_title":"Defending Jailbreak Attacks on Large Language Models via Manifold Trajectory Kinetics","ref_index":90,"is_internal_anchor":false},{"citing_arxiv_id":"2606.07631","citing_title":"Trait-space Monitoring for Emergent Misalignment During Supervised Finetuning","ref_index":63,"is_internal_anchor":false},{"citing_arxiv_id":"2605.26315","citing_title":"Curriculum Learning for Safety Alignment","ref_index":10,"is_internal_anchor":false},{"citing_arxiv_id":"2605.29659","citing_title":"Opir: Efficient Multi-Task Safety Classification for Toxicity, Jailbreaks, Hate Speech, and Harmful Content","ref_index":22,"is_internal_anchor":false},{"citing_arxiv_id":"2502.06387","citing_title":"How Humans Help LLMs: Assessing and Incentivizing Human Preference Annotators","ref_index":53,"is_internal_anchor":false},{"citing_arxiv_id":"2505.19134","citing_title":"Incentivizing High-Quality Human Annotations with Golden Questions","ref_index":23,"is_internal_anchor":false},{"citing_arxiv_id":"2511.02623","citing_title":"The Realignment Problem: When Right becomes Wrong in LLMs","ref_index":10,"is_internal_anchor":false},{"citing_arxiv_id":"2512.10998","citing_title":"SCOUT: A Defense Against Data Poisoning Attacks in Fine-Tuned Language Models","ref_index":29,"is_internal_anchor":false},{"citing_arxiv_id":"2602.07892","citing_title":"Safety Alignment as Continual Learning: Mitigating the Alignment Tax via Orthogonal Gradient Projection","ref_index":14,"is_internal_anchor":false},{"citing_arxiv_id":"2605.09946","citing_title":"Structure from Strategic Interaction & Uncertainty: Risk Sensitive Games for Robust Preference Learning","ref_index":69,"is_internal_anchor":false},{"citing_arxiv_id":"2605.11679","citing_title":"Explaining and Breaking the Safety-Helpfulness Ceiling via Preference Dimensional Expansion","ref_index":22,"is_internal_anchor":false},{"citing_arxiv_id":"2605.11712","citing_title":"Toward Stable Value Alignment: Introducing Independent Modules for Consistent Value Guidance","ref_index":81,"is_internal_anchor":false},{"citing_arxiv_id":"2605.11679","citing_title":"Explaining and Breaking the Safety-Helpfulness Ceiling via Preference Dimensional Expansion","ref_index":22,"is_internal_anchor":false},{"citing_arxiv_id":"2605.09946","citing_title":"Structure from Strategic Interaction & Uncertainty: Risk Sensitive Games for Robust Preference Learning","ref_index":69,"is_internal_anchor":false},{"citing_arxiv_id":"2412.05579","citing_title":"LLMs-as-Judges: A Comprehensive Survey on LLM-based Evaluation Methods","ref_index":100,"is_internal_anchor":false},{"citing_arxiv_id":"2605.01899","citing_title":"Disentangling Intent from Role: Adversarial Self-Play for Persona-Invariant Safety Alignment","ref_index":46,"is_internal_anchor":false},{"citing_arxiv_id":"2604.10673","citing_title":"Principles Do Not Apply Themselves: A Hermeneutic Perspective on AI Alignment","ref_index":13,"is_internal_anchor":false},{"citing_arxiv_id":"2604.06833","citing_title":"FedDetox: Robust Federated SLM Alignment via On-Device Data Sanitization","ref_index":25,"is_internal_anchor":false},{"citing_arxiv_id":"2605.07105","citing_title":"Theoretical Limits of Language Model Alignment","ref_index":28,"is_internal_anchor":false},{"citing_arxiv_id":"2605.07982","citing_title":"GLiGuard: Schema-Conditioned Classification for LLM Safeguard","ref_index":5,"is_internal_anchor":false},{"citing_arxiv_id":"2604.18519","citing_title":"LLM Safety From Within: Detecting Harmful Content with Internal Representations","ref_index":8,"is_internal_anchor":false},{"citing_arxiv_id":"2604.17614","citing_title":"Characterizing Model-Native Skills","ref_index":99,"is_internal_anchor":false}]},"formal_canon":{"evidence_count":0,"sample":[],"anchors":[]},"links":{"html":"https://pith.science/pith/DIFWTZLUGIWN5BBP5AMRODX7Y6","json":"https://pith.science/pith/DIFWTZLUGIWN5BBP5AMRODX7Y6.json","graph_json":"https://pith.science/api/pith-number/DIFWTZLUGIWN5BBP5AMRODX7Y6/graph.json","events_json":"https://pith.science/api/pith-number/DIFWTZLUGIWN5BBP5AMRODX7Y6/events.json","paper":"https://pith.science/paper/DIFWTZLU"},"agent_actions":{"view_html":"https://pith.science/pith/DIFWTZLUGIWN5BBP5AMRODX7Y6","download_json":"https://pith.science/pith/DIFWTZLUGIWN5BBP5AMRODX7Y6.json","view_paper":"https://pith.science/paper/DIFWTZLU","resolve_alias":"https://pith.science/api/pith-number/resolve?arxiv=2406.15513&json=true","fetch_graph":"https://pith.science/api/pith-number/DIFWTZLUGIWN5BBP5AMRODX7Y6/graph.json","fetch_events":"https://pith.science/api/pith-number/DIFWTZLUGIWN5BBP5AMRODX7Y6/events.json","actions":{"anchor_timestamp":"https://pith.science/pith/DIFWTZLUGIWN5BBP5AMRODX7Y6/action/timestamp_anchor","attest_storage":"https://pith.science/pith/DIFWTZLUGIWN5BBP5AMRODX7Y6/action/storage_attestation","attest_author":"https://pith.science/pith/DIFWTZLUGIWN5BBP5AMRODX7Y6/action/author_attestation","sign_citation":"https://pith.science/pith/DIFWTZLUGIWN5BBP5AMRODX7Y6/action/citation_signature","submit_replication":"https://pith.science/pith/DIFWTZLUGIWN5BBP5AMRODX7Y6/action/replication_record"}},"created_at":"2026-07-05T11:21:33.639010+00:00","updated_at":"2026-07-05T11:21:33.639010+00:00"}