{"record_type":"pith_number_record","schema_url":"https://pith.science/schemas/pith-number/v1.json","pith_number":"pith:2023:BLIM427O5PBWNVX4JI2MLIWVSU","short_pith_number":"pith:BLIM427O","schema_version":"1.0","canonical_sha256":"0ad0ce6beeebc366d6fc4a34c5a2d5953eb85748c6cd52e89a43c47b6e39d006","source":{"kind":"arxiv","id":"2310.02238","version":2},"attestation_state":"computed","paper":{"title":"Who's Harry Potter? Approximate Unlearning in LLMs","license":"http://creativecommons.org/licenses/by/4.0/","headline":"","cross_cats":["cs.AI"],"primary_cat":"cs.CL","authors_text":"Mark Russinovich, Ronen Eldan","submitted_at":"2023-10-03T17:48:14Z","abstract_excerpt":"Large language models (LLMs) are trained on massive internet corpora that often contain copyrighted content. This poses legal and ethical challenges for the developers and users of these models, as well as the original authors and publishers. In this paper, we propose a novel technique for unlearning a subset of the training data from a LLM, without having to retrain it from scratch.\n  We evaluate our technique on the task of unlearning the Harry Potter books from the Llama2-7b model (a generative language model recently open-sourced by Meta). While the model took over 184K GPU-hours to pretra"},"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":"2310.02238","kind":"arxiv","version":2},"metadata":{"license":"http://creativecommons.org/licenses/by/4.0/","primary_cat":"cs.CL","submitted_at":"2023-10-03T17:48:14Z","cross_cats_sorted":["cs.AI"],"title_canon_sha256":"44472a32176e1dca805c9f34ea4e0d43da9bc06d7be869d069d1fe4520a844f2","abstract_canon_sha256":"7b72a6af43d6680ad48aca4c2954df2a88ec4558cc64c7951f69353c58d21477"},"schema_version":"1.0"},"receipt":{"kind":"pith_receipt","key_id":"pith-v1-2026-05","algorithm":"ed25519","signed_at":"2026-07-05T06:57:11.521572Z","signature_b64":"WjWQi/dT24iXz8jU1FuAZD2uX3K21e0t92+l4Wh5VPFnJ8jAzWdGUEbZaqZ7VVBbZcPAiGEfvIOre4n3Bt9OCQ==","signed_message":"canonical_sha256_bytes","builder_version":"pith-number-builder-2026-05-17-v1","receipt_version":"0.3","canonical_sha256":"0ad0ce6beeebc366d6fc4a34c5a2d5953eb85748c6cd52e89a43c47b6e39d006","last_reissued_at":"2026-07-05T06:57:11.521126Z","signature_status":"signed_v1","first_computed_at":"2026-07-05T06:57:11.521126Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"graph_snapshot":{"paper":{"title":"Who's Harry Potter? Approximate Unlearning in LLMs","license":"http://creativecommons.org/licenses/by/4.0/","headline":"","cross_cats":["cs.AI"],"primary_cat":"cs.CL","authors_text":"Mark Russinovich, Ronen Eldan","submitted_at":"2023-10-03T17:48:14Z","abstract_excerpt":"Large language models (LLMs) are trained on massive internet corpora that often contain copyrighted content. This poses legal and ethical challenges for the developers and users of these models, as well as the original authors and publishers. In this paper, we propose a novel technique for unlearning a subset of the training data from a LLM, without having to retrain it from scratch.\n  We evaluate our technique on the task of unlearning the Harry Potter books from the Llama2-7b model (a generative language model recently open-sourced by Meta). While the model took over 184K GPU-hours to pretra"},"claims":{"count":0,"items":[],"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"source":{"id":"2310.02238","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/2310.02238/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":"2310.02238","created_at":"2026-07-05T06:57:11.521186+00:00"},{"alias_kind":"arxiv_version","alias_value":"2310.02238v2","created_at":"2026-07-05T06:57:11.521186+00:00"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2310.02238","created_at":"2026-07-05T06:57:11.521186+00:00"},{"alias_kind":"pith_short_12","alias_value":"BLIM427O5PBW","created_at":"2026-07-05T06:57:11.521186+00:00"},{"alias_kind":"pith_short_16","alias_value":"BLIM427O5PBWNVX4","created_at":"2026-07-05T06:57:11.521186+00:00"},{"alias_kind":"pith_short_8","alias_value":"BLIM427O","created_at":"2026-07-05T06:57:11.521186+00:00"}],"events":[],"event_summary":{},"paper_claims":[],"inbound_citations":{"count":37,"internal_anchor_count":0,"sample":[{"citing_arxiv_id":"2607.02513","citing_title":"LACUNA: A Testbed for Evaluating Localization Precision for LLM Unlearning","ref_index":29,"is_internal_anchor":false},{"citing_arxiv_id":"2606.12841","citing_title":"TimeROME-DLM: Temporal Causal Tracing and Low-Rank Inference-Time Knowledge Editing for Masked Diffusion Language Models","ref_index":31,"is_internal_anchor":false},{"citing_arxiv_id":"2606.10989","citing_title":"Null-Space Constrained Low-Rank Adaptation for Response-Specified Large Language Model Unlearning","ref_index":17,"is_internal_anchor":false},{"citing_arxiv_id":"2606.07688","citing_title":"TRACER: Token ReAssignment for Concept ERasure in Generative Recommendation","ref_index":17,"is_internal_anchor":false},{"citing_arxiv_id":"2606.07141","citing_title":"REMEDI: A Benchmark for Retention and Unlearning Evaluation in Multi-label Clinical Disease Inference","ref_index":12,"is_internal_anchor":false},{"citing_arxiv_id":"2606.04182","citing_title":"Exact Unlearning in Reinforcement Learning","ref_index":31,"is_internal_anchor":false},{"citing_arxiv_id":"2606.02453","citing_title":"Initialization is Half the Battle: Generating Diverse Images from a Guidance Potential Posterior","ref_index":48,"is_internal_anchor":false},{"citing_arxiv_id":"2606.02920","citing_title":"Fast Unlearning at Scale via Margin Self-Correction","ref_index":12,"is_internal_anchor":false},{"citing_arxiv_id":"2606.02119","citing_title":"How Hard Can It Be? Hardness-Aware Multi-Objective Unlearning","ref_index":2,"is_internal_anchor":false},{"citing_arxiv_id":"2605.01129","citing_title":"Revisiting Privacy Leakage in Machine Unlearning: Membership Inference Beyond the Forgotten Set","ref_index":67,"is_internal_anchor":false},{"citing_arxiv_id":"2606.30788","citing_title":"Revocable Learned State via Process Sidecars","ref_index":4,"is_internal_anchor":false},{"citing_arxiv_id":"2605.18879","citing_title":"ZeroUnlearn: Few-Shot Knowledge Unlearning in Large Language Models","ref_index":3,"is_internal_anchor":false},{"citing_arxiv_id":"2605.26133","citing_title":"Pretraining Data Exposure in Large Language Models: A Survey of Membership Inference, Data Contamination, and Security Implications","ref_index":14,"is_internal_anchor":false},{"citing_arxiv_id":"2606.00105","citing_title":"Visual-Noise Guided In-Context Distillation for Multimodal Large Language Model Unlearning","ref_index":40,"is_internal_anchor":false},{"citing_arxiv_id":"2606.02293","citing_title":"AI as a Tool for Simulation-Based Experiments in Literary Studies","ref_index":16,"is_internal_anchor":false},{"citing_arxiv_id":"2408.12935","citing_title":"AI Safety Landscape for Large Language Models: Taxonomy, State-of-the-art, and Future Directions","ref_index":197,"is_internal_anchor":false},{"citing_arxiv_id":"2501.19202","citing_title":"Improving LLM Unlearning Robustness via Random Perturbations","ref_index":10,"is_internal_anchor":false},{"citing_arxiv_id":"2605.18879","citing_title":"ZeroUnlearn: Few-Shot Knowledge Unlearning in Large Language Models","ref_index":3,"is_internal_anchor":false},{"citing_arxiv_id":"2605.20915","citing_title":"Calibration vs Decision Making: Revisiting the Reliability Paradox in Unlearned Language Models","ref_index":45,"is_internal_anchor":false},{"citing_arxiv_id":"2605.18879","citing_title":"ZeroUnlearn: Few-Shot Knowledge Unlearning in Large Language Models","ref_index":3,"is_internal_anchor":false},{"citing_arxiv_id":"2605.18891","citing_title":"Auditing Reasoning-Trace Memorization Claims after Unlearning with Head-Conditioned Canaries","ref_index":4,"is_internal_anchor":false},{"citing_arxiv_id":"2506.20941","citing_title":"Revisiting the Past: Data Unlearning with Model State History","ref_index":12,"is_internal_anchor":false},{"citing_arxiv_id":"2509.22483","citing_title":"OFMU: Optimization-Driven Framework for Machine Unlearning","ref_index":8,"is_internal_anchor":false},{"citing_arxiv_id":"2511.02623","citing_title":"The Realignment Problem: When Right becomes Wrong in LLMs","ref_index":6,"is_internal_anchor":false},{"citing_arxiv_id":"2310.16789","citing_title":"Detecting Pretraining Data from Large Language Models","ref_index":14,"is_internal_anchor":false}]},"formal_canon":{"evidence_count":0,"sample":[],"anchors":[]},"links":{"html":"https://pith.science/pith/BLIM427O5PBWNVX4JI2MLIWVSU","json":"https://pith.science/pith/BLIM427O5PBWNVX4JI2MLIWVSU.json","graph_json":"https://pith.science/api/pith-number/BLIM427O5PBWNVX4JI2MLIWVSU/graph.json","events_json":"https://pith.science/api/pith-number/BLIM427O5PBWNVX4JI2MLIWVSU/events.json","paper":"https://pith.science/paper/BLIM427O"},"agent_actions":{"view_html":"https://pith.science/pith/BLIM427O5PBWNVX4JI2MLIWVSU","download_json":"https://pith.science/pith/BLIM427O5PBWNVX4JI2MLIWVSU.json","view_paper":"https://pith.science/paper/BLIM427O","resolve_alias":"https://pith.science/api/pith-number/resolve?arxiv=2310.02238&json=true","fetch_graph":"https://pith.science/api/pith-number/BLIM427O5PBWNVX4JI2MLIWVSU/graph.json","fetch_events":"https://pith.science/api/pith-number/BLIM427O5PBWNVX4JI2MLIWVSU/events.json","actions":{"anchor_timestamp":"https://pith.science/pith/BLIM427O5PBWNVX4JI2MLIWVSU/action/timestamp_anchor","attest_storage":"https://pith.science/pith/BLIM427O5PBWNVX4JI2MLIWVSU/action/storage_attestation","attest_author":"https://pith.science/pith/BLIM427O5PBWNVX4JI2MLIWVSU/action/author_attestation","sign_citation":"https://pith.science/pith/BLIM427O5PBWNVX4JI2MLIWVSU/action/citation_signature","submit_replication":"https://pith.science/pith/BLIM427O5PBWNVX4JI2MLIWVSU/action/replication_record"}},"created_at":"2026-07-05T06:57:11.521186+00:00","updated_at":"2026-07-05T06:57:11.521186+00:00"}