{"record_type":"pith_number_record","schema_url":"https://pith.science/schemas/pith-number/v1.json","pith_number":"pith:2023:TGPGAX7WJSZNR7A4QYFKZC4BGY","short_pith_number":"pith:TGPGAX7W","schema_version":"1.0","canonical_sha256":"999e605ff64cb2d8fc1c860aac8b813639355d6b302a81e1c6b9e9a89a3eac76","source":{"kind":"arxiv","id":"2310.20624","version":2},"attestation_state":"computed","paper":{"title":"LoRA Fine-tuning Efficiently Undoes Safety Training in Llama 2-Chat 70B","license":"http://creativecommons.org/licenses/by/4.0/","headline":"","cross_cats":["cs.AI"],"primary_cat":"cs.LG","authors_text":"Charlie Rogers-Smith, Jeffrey Ladish, Simon Lermen","submitted_at":"2023-10-31T16:55:06Z","abstract_excerpt":"AI developers often apply safety alignment procedures to prevent the misuse of their AI systems. For example, before Meta released Llama 2-Chat - a collection of instruction fine-tuned large language models - they invested heavily in safety training, incorporating extensive red-teaming and reinforcement learning from human feedback. We explore the robustness of safety training in language models by subversively fine-tuning Llama 2-Chat. We employ quantized low-rank adaptation (LoRA) as an efficient fine-tuning method. With a budget of less than \\$200 and using only one GPU, we successfully und"},"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":"2310.20624","kind":"arxiv","version":2},"metadata":{"license":"http://creativecommons.org/licenses/by/4.0/","primary_cat":"cs.LG","submitted_at":"2023-10-31T16:55:06Z","cross_cats_sorted":["cs.AI"],"title_canon_sha256":"8b0e5c1df72f88a5475a4ef8da559e4d21e83ea7b5bec580fa31cf433422efd1","abstract_canon_sha256":"2c313efb5794d4d4071056045b3008be2ba12e5cab529e9401ed9b82c5957eb7"},"schema_version":"1.0"},"receipt":{"kind":"pith_receipt","key_id":"pith-v1-2026-05","algorithm":"ed25519","signed_at":"2026-07-05T08:21:42.728877Z","signature_b64":"CSYtiamAoFtt63Ws2hVsrLSBFmVOTeMIjU8ftgjOOykpPU04/uipAs3gaLr7NqqeMNn769NXdjhnnWwXy9okBg==","signed_message":"canonical_sha256_bytes","builder_version":"pith-number-builder-2026-05-17-v1","receipt_version":"0.3","canonical_sha256":"999e605ff64cb2d8fc1c860aac8b813639355d6b302a81e1c6b9e9a89a3eac76","last_reissued_at":"2026-07-05T08:21:42.728380Z","signature_status":"signed_v1","first_computed_at":"2026-07-05T08:21:42.728380Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"graph_snapshot":{"paper":{"title":"LoRA Fine-tuning Efficiently Undoes Safety Training in Llama 2-Chat 70B","license":"http://creativecommons.org/licenses/by/4.0/","headline":"","cross_cats":["cs.AI"],"primary_cat":"cs.LG","authors_text":"Charlie Rogers-Smith, Jeffrey Ladish, Simon Lermen","submitted_at":"2023-10-31T16:55:06Z","abstract_excerpt":"AI developers often apply safety alignment procedures to prevent the misuse of their AI systems. For example, before Meta released Llama 2-Chat - a collection of instruction fine-tuned large language models - they invested heavily in safety training, incorporating extensive red-teaming and reinforcement learning from human feedback. We explore the robustness of safety training in language models by subversively fine-tuning Llama 2-Chat. We employ quantized low-rank adaptation (LoRA) as an efficient fine-tuning method. With a budget of less than \\$200 and using only one GPU, we successfully und"},"claims":{"count":0,"items":[],"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"source":{"id":"2310.20624","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/2310.20624/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":"2310.20624","created_at":"2026-07-05T08:21:42.728440+00:00"},{"alias_kind":"arxiv_version","alias_value":"2310.20624v2","created_at":"2026-07-05T08:21:42.728440+00:00"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2310.20624","created_at":"2026-07-05T08:21:42.728440+00:00"},{"alias_kind":"pith_short_12","alias_value":"TGPGAX7WJSZN","created_at":"2026-07-05T08:21:42.728440+00:00"},{"alias_kind":"pith_short_16","alias_value":"TGPGAX7WJSZNR7A4","created_at":"2026-07-05T08:21:42.728440+00:00"},{"alias_kind":"pith_short_8","alias_value":"TGPGAX7W","created_at":"2026-07-05T08:21:42.728440+00:00"}],"events":[],"event_summary":{},"paper_claims":[],"inbound_citations":{"count":26,"internal_anchor_count":0,"sample":[{"citing_arxiv_id":"2606.26620","citing_title":"Discovering Millions of Interpretable Features with Sparse Autoencoders","ref_index":29,"is_internal_anchor":false},{"citing_arxiv_id":"2606.17168","citing_title":"RepSelect: Robust LLM Unlearning via Representation Selectivity","ref_index":22,"is_internal_anchor":false},{"citing_arxiv_id":"2606.07631","citing_title":"Trait-space Monitoring for Emergent Misalignment During Supervised Finetuning","ref_index":5,"is_internal_anchor":false},{"citing_arxiv_id":"2606.00160","citing_title":"DataShield: Safety-degrading Data Filtering for LLM Benign Instruction Fine-Tuning","ref_index":4,"is_internal_anchor":false},{"citing_arxiv_id":"2605.12850","citing_title":"Persona-Model Collapse in Emergent Misalignment","ref_index":35,"is_internal_anchor":false},{"citing_arxiv_id":"2605.14605","citing_title":"One Step to the Side: Why Defenses Against Malicious Finetuning Fail Under Adaptive Adversaries","ref_index":19,"is_internal_anchor":false},{"citing_arxiv_id":"2605.24583","citing_title":"Measuring Alignment-Induced Activation Shifts Correctly: A Template-Controlled Difference-in-Differences Protocol","ref_index":9,"is_internal_anchor":false},{"citing_arxiv_id":"2605.24550","citing_title":"Jailbreak to Protect: Buffering and Reinforcing via Temporary Jailbreaking for Safe Fine-Tuning in Large Language Models","ref_index":27,"is_internal_anchor":false},{"citing_arxiv_id":"2606.28843","citing_title":"The Heterogeneous Safety Impacts of Benign Multilingual Fine-Tuning","ref_index":15,"is_internal_anchor":false},{"citing_arxiv_id":"2605.25073","citing_title":"Security in the Fine-Tuning Lifecycle of Large Language Models: Threats, Defenses,Evaluation, and Future Directions","ref_index":82,"is_internal_anchor":false},{"citing_arxiv_id":"2605.26315","citing_title":"Curriculum Learning for Safety Alignment","ref_index":12,"is_internal_anchor":false},{"citing_arxiv_id":"2605.26526","citing_title":"Open-Weight LLM Fine-Tuning Defenses are Susceptible to Simple Attacks","ref_index":7,"is_internal_anchor":false},{"citing_arxiv_id":"2605.28030","citing_title":"SPARD: Defending Harmful Fine-Tuning Attack via Safety Projection with Relevance-Diversity Data Selection","ref_index":8,"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":85,"is_internal_anchor":false},{"citing_arxiv_id":"2605.21674","citing_title":"Adversarial Reframing: A Framework for Targeted Generation in Language Models","ref_index":28,"is_internal_anchor":false},{"citing_arxiv_id":"2508.20697","citing_title":"Token Buncher: Shielding LLMs from Harmful Reinforcement Learning Fine-Tuning","ref_index":26,"is_internal_anchor":false},{"citing_arxiv_id":"2407.04295","citing_title":"Jailbreak Attacks and Defenses Against Large Language Models: A Survey","ref_index":47,"is_internal_anchor":false},{"citing_arxiv_id":"2605.12850","citing_title":"Persona-Model Collapse in Emergent Misalignment","ref_index":35,"is_internal_anchor":false},{"citing_arxiv_id":"2604.02574","citing_title":"Understanding the Effects of Safety Unalignment on Large Language Models","ref_index":20,"is_internal_anchor":false},{"citing_arxiv_id":"2406.11717","citing_title":"Refusal in Language Models Is Mediated by a Single Direction","ref_index":146,"is_internal_anchor":false},{"citing_arxiv_id":"2605.10998","citing_title":"Few-Shot Truly Benign DPO Attack for Jailbreaking LLMs","ref_index":19,"is_internal_anchor":false},{"citing_arxiv_id":"2604.24902","citing_title":"Safety Drift After Fine-Tuning: Evidence from High-Stakes Domains","ref_index":32,"is_internal_anchor":false},{"citing_arxiv_id":"2605.04992","citing_title":"You Snooze, You Lose: Automatic Safety Alignment Restoration through Neural Weight Translation","ref_index":79,"is_internal_anchor":false},{"citing_arxiv_id":"2604.21905","citing_title":"Low-Rank Adaptation Redux for Large Models","ref_index":100,"is_internal_anchor":false},{"citing_arxiv_id":"2604.18510","citing_title":"Different Paths to Harmful Compliance: Behavioral Side Effects and Mechanistic Divergence Across LLM Jailbreaks","ref_index":39,"is_internal_anchor":false}]},"formal_canon":{"evidence_count":0,"sample":[],"anchors":[]},"links":{"html":"https://pith.science/pith/TGPGAX7WJSZNR7A4QYFKZC4BGY","json":"https://pith.science/pith/TGPGAX7WJSZNR7A4QYFKZC4BGY.json","graph_json":"https://pith.science/api/pith-number/TGPGAX7WJSZNR7A4QYFKZC4BGY/graph.json","events_json":"https://pith.science/api/pith-number/TGPGAX7WJSZNR7A4QYFKZC4BGY/events.json","paper":"https://pith.science/paper/TGPGAX7W"},"agent_actions":{"view_html":"https://pith.science/pith/TGPGAX7WJSZNR7A4QYFKZC4BGY","download_json":"https://pith.science/pith/TGPGAX7WJSZNR7A4QYFKZC4BGY.json","view_paper":"https://pith.science/paper/TGPGAX7W","resolve_alias":"https://pith.science/api/pith-number/resolve?arxiv=2310.20624&json=true","fetch_graph":"https://pith.science/api/pith-number/TGPGAX7WJSZNR7A4QYFKZC4BGY/graph.json","fetch_events":"https://pith.science/api/pith-number/TGPGAX7WJSZNR7A4QYFKZC4BGY/events.json","actions":{"anchor_timestamp":"https://pith.science/pith/TGPGAX7WJSZNR7A4QYFKZC4BGY/action/timestamp_anchor","attest_storage":"https://pith.science/pith/TGPGAX7WJSZNR7A4QYFKZC4BGY/action/storage_attestation","attest_author":"https://pith.science/pith/TGPGAX7WJSZNR7A4QYFKZC4BGY/action/author_attestation","sign_citation":"https://pith.science/pith/TGPGAX7WJSZNR7A4QYFKZC4BGY/action/citation_signature","submit_replication":"https://pith.science/pith/TGPGAX7WJSZNR7A4QYFKZC4BGY/action/replication_record"}},"created_at":"2026-07-05T08:21:42.728440+00:00","updated_at":"2026-07-05T08:21:42.728440+00:00"}