{"record_type":"pith_number_record","schema_url":"https://pith.science/schemas/pith-number/v1.json","pith_number":"pith:2024:754W2V7D5TPV2TAH2L36LCB3FG","short_pith_number":"pith:754W2V7D","schema_version":"1.0","canonical_sha256":"ff796d57e3ecdf5d4c07d2f7e5883b29be6b88a97ff0ea4410ef3eec3bbb912b","source":{"kind":"arxiv","id":"2402.10208","version":2},"attestation_state":"computed","paper":{"title":"Recovering the Pre-Fine-Tuning Weights of Generative Models","license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","headline":"","cross_cats":["cs.CL","cs.CR","cs.CV"],"primary_cat":"cs.LG","authors_text":"Eliahu Horwitz, Jonathan Kahana, Yedid Hoshen","submitted_at":"2024-02-15T18:59:02Z","abstract_excerpt":"The dominant paradigm in generative modeling consists of two steps: i) pre-training on a large-scale but unsafe dataset, ii) aligning the pre-trained model with human values via fine-tuning. This practice is considered safe, as no current method can recover the unsafe, pre-fine-tuning model weights. In this paper, we demonstrate that this assumption is often false. Concretely, we present Spectral DeTuning, a method that can recover the weights of the pre-fine-tuning model using a few low-rank (LoRA) fine-tuned models. In contrast to previous attacks that attempt to recover pre-fine-tuning capa"},"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":"2402.10208","kind":"arxiv","version":2},"metadata":{"license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","primary_cat":"cs.LG","submitted_at":"2024-02-15T18:59:02Z","cross_cats_sorted":["cs.CL","cs.CR","cs.CV"],"title_canon_sha256":"ffbb2ffa43297e2add9cc761c691074d0be8edf878450da1ec702e215b7d9a4d","abstract_canon_sha256":"2b71715fd2796b2bb9e29f95eaac0593e70d84fccdf602b9c70cbb4a90537225"},"schema_version":"1.0"},"receipt":{"kind":"pith_receipt","key_id":"pith-v1-2026-05","algorithm":"ed25519","signed_at":"2026-07-05T08:38:36.433707Z","signature_b64":"3GD9vj+nK0QQRxn7Inmmk9aCgKNA+eRgFxC+WEalOjFMI3EazxFYFtGl6YC+cmP5YJcmhJyfgPML43CciHYXCA==","signed_message":"canonical_sha256_bytes","builder_version":"pith-number-builder-2026-05-17-v1","receipt_version":"0.3","canonical_sha256":"ff796d57e3ecdf5d4c07d2f7e5883b29be6b88a97ff0ea4410ef3eec3bbb912b","last_reissued_at":"2026-07-05T08:38:36.433221Z","signature_status":"signed_v1","first_computed_at":"2026-07-05T08:38:36.433221Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"graph_snapshot":{"paper":{"title":"Recovering the Pre-Fine-Tuning Weights of Generative Models","license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","headline":"","cross_cats":["cs.CL","cs.CR","cs.CV"],"primary_cat":"cs.LG","authors_text":"Eliahu Horwitz, Jonathan Kahana, Yedid Hoshen","submitted_at":"2024-02-15T18:59:02Z","abstract_excerpt":"The dominant paradigm in generative modeling consists of two steps: i) pre-training on a large-scale but unsafe dataset, ii) aligning the pre-trained model with human values via fine-tuning. This practice is considered safe, as no current method can recover the unsafe, pre-fine-tuning model weights. In this paper, we demonstrate that this assumption is often false. Concretely, we present Spectral DeTuning, a method that can recover the weights of the pre-fine-tuning model using a few low-rank (LoRA) fine-tuned models. In contrast to previous attacks that attempt to recover pre-fine-tuning capa"},"claims":{"count":0,"items":[],"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"source":{"id":"2402.10208","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/2402.10208/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":"2402.10208","created_at":"2026-07-05T08:38:36.433279+00:00"},{"alias_kind":"arxiv_version","alias_value":"2402.10208v2","created_at":"2026-07-05T08:38:36.433279+00:00"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2402.10208","created_at":"2026-07-05T08:38:36.433279+00:00"},{"alias_kind":"pith_short_12","alias_value":"754W2V7D5TPV","created_at":"2026-07-05T08:38:36.433279+00:00"},{"alias_kind":"pith_short_16","alias_value":"754W2V7D5TPV2TAH","created_at":"2026-07-05T08:38:36.433279+00:00"},{"alias_kind":"pith_short_8","alias_value":"754W2V7D","created_at":"2026-07-05T08:38:36.433279+00:00"}],"events":[],"event_summary":{},"paper_claims":[],"inbound_citations":{"count":1,"internal_anchor_count":1,"sample":[{"citing_arxiv_id":"2502.00840","citing_title":"Activation Approximations Can Incur Safety Vulnerabilities Even in Aligned LLMs: Comprehensive Analysis and Defense","ref_index":63,"is_internal_anchor":true}]},"formal_canon":{"evidence_count":0,"sample":[],"anchors":[]},"links":{"html":"https://pith.science/pith/754W2V7D5TPV2TAH2L36LCB3FG","json":"https://pith.science/pith/754W2V7D5TPV2TAH2L36LCB3FG.json","graph_json":"https://pith.science/api/pith-number/754W2V7D5TPV2TAH2L36LCB3FG/graph.json","events_json":"https://pith.science/api/pith-number/754W2V7D5TPV2TAH2L36LCB3FG/events.json","paper":"https://pith.science/paper/754W2V7D"},"agent_actions":{"view_html":"https://pith.science/pith/754W2V7D5TPV2TAH2L36LCB3FG","download_json":"https://pith.science/pith/754W2V7D5TPV2TAH2L36LCB3FG.json","view_paper":"https://pith.science/paper/754W2V7D","resolve_alias":"https://pith.science/api/pith-number/resolve?arxiv=2402.10208&json=true","fetch_graph":"https://pith.science/api/pith-number/754W2V7D5TPV2TAH2L36LCB3FG/graph.json","fetch_events":"https://pith.science/api/pith-number/754W2V7D5TPV2TAH2L36LCB3FG/events.json","actions":{"anchor_timestamp":"https://pith.science/pith/754W2V7D5TPV2TAH2L36LCB3FG/action/timestamp_anchor","attest_storage":"https://pith.science/pith/754W2V7D5TPV2TAH2L36LCB3FG/action/storage_attestation","attest_author":"https://pith.science/pith/754W2V7D5TPV2TAH2L36LCB3FG/action/author_attestation","sign_citation":"https://pith.science/pith/754W2V7D5TPV2TAH2L36LCB3FG/action/citation_signature","submit_replication":"https://pith.science/pith/754W2V7D5TPV2TAH2L36LCB3FG/action/replication_record"}},"created_at":"2026-07-05T08:38:36.433279+00:00","updated_at":"2026-07-05T08:38:36.433279+00:00"}