{"record_type":"pith_number_record","schema_url":"https://pith.science/schemas/pith-number/v1.json","pith_number":"pith:2024:73CMOHP6Z6RAFZDXDV54L7LJKE","short_pith_number":"pith:73CMOHP6","schema_version":"1.0","canonical_sha256":"fec4c71dfecfa202e4771d7bc5fd69512d1df8507d90f9890f5f34fa64389cb6","source":{"kind":"arxiv","id":"2410.15821","version":1},"attestation_state":"computed","paper":{"title":"The effect of fine-tuning on language model toxicity","license":"http://creativecommons.org/licenses/by/4.0/","headline":"","cross_cats":[],"primary_cat":"cs.AI","authors_text":"Brent Mittelstadt, Chris Russell, Will Hawkins","submitted_at":"2024-10-21T09:39:09Z","abstract_excerpt":"Fine-tuning language models has become increasingly popular following the proliferation of open models and improvements in cost-effective parameter efficient fine-tuning. However, fine-tuning can influence model properties such as safety. We assess how fine-tuning can impact different open models' propensity to output toxic content. We assess the impacts of fine-tuning Gemma, Llama, and Phi models on toxicity through three experiments. We compare how toxicity is reduced by model developers during instruction-tuning. We show that small amounts of parameter-efficient fine-tuning on developer-tun"},"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":"2410.15821","kind":"arxiv","version":1},"metadata":{"license":"http://creativecommons.org/licenses/by/4.0/","primary_cat":"cs.AI","submitted_at":"2024-10-21T09:39:09Z","cross_cats_sorted":[],"title_canon_sha256":"afa55db7ae212b45409a45394b9cbd313fefcdc9732285eead37a71827e5201c","abstract_canon_sha256":"5368d3060a7b2c70a22194ce6c2f901fffcb0f5f01c45893bb997e7d01f4ec78"},"schema_version":"1.0"},"receipt":{"kind":"pith_receipt","key_id":"pith-v1-2026-05","algorithm":"ed25519","signed_at":"2026-07-05T09:23:25.377467Z","signature_b64":"rNZFpKbNaiN2V8jgb9mISUJ0ddjE6lw7AWJOyEE4CT2s+V60WREFC8BJjmKc+6fUbo+lWs945G+DPtjamcmHBw==","signed_message":"canonical_sha256_bytes","builder_version":"pith-number-builder-2026-05-17-v1","receipt_version":"0.3","canonical_sha256":"fec4c71dfecfa202e4771d7bc5fd69512d1df8507d90f9890f5f34fa64389cb6","last_reissued_at":"2026-07-05T09:23:25.377038Z","signature_status":"signed_v1","first_computed_at":"2026-07-05T09:23:25.377038Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"graph_snapshot":{"paper":{"title":"The effect of fine-tuning on language model toxicity","license":"http://creativecommons.org/licenses/by/4.0/","headline":"","cross_cats":[],"primary_cat":"cs.AI","authors_text":"Brent Mittelstadt, Chris Russell, Will Hawkins","submitted_at":"2024-10-21T09:39:09Z","abstract_excerpt":"Fine-tuning language models has become increasingly popular following the proliferation of open models and improvements in cost-effective parameter efficient fine-tuning. However, fine-tuning can influence model properties such as safety. We assess how fine-tuning can impact different open models' propensity to output toxic content. We assess the impacts of fine-tuning Gemma, Llama, and Phi models on toxicity through three experiments. We compare how toxicity is reduced by model developers during instruction-tuning. We show that small amounts of parameter-efficient fine-tuning on developer-tun"},"claims":{"count":0,"items":[],"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"source":{"id":"2410.15821","kind":"arxiv","version":1},"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/2410.15821/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":"2410.15821","created_at":"2026-07-05T09:23:25.377099+00:00"},{"alias_kind":"arxiv_version","alias_value":"2410.15821v1","created_at":"2026-07-05T09:23:25.377099+00:00"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2410.15821","created_at":"2026-07-05T09:23:25.377099+00:00"},{"alias_kind":"pith_short_12","alias_value":"73CMOHP6Z6RA","created_at":"2026-07-05T09:23:25.377099+00:00"},{"alias_kind":"pith_short_16","alias_value":"73CMOHP6Z6RAFZDX","created_at":"2026-07-05T09:23:25.377099+00:00"},{"alias_kind":"pith_short_8","alias_value":"73CMOHP6","created_at":"2026-07-05T09:23:25.377099+00:00"}],"events":[],"event_summary":{},"paper_claims":[],"inbound_citations":{"count":5,"internal_anchor_count":0,"sample":[{"citing_arxiv_id":"2606.25380","citing_title":"A Survey of Toxicity Detection and Mitigation Strategies for Multilingual Language Models","ref_index":117,"is_internal_anchor":false},{"citing_arxiv_id":"2606.30263","citing_title":"Defending Against Harmful Supervision Hidden in Benign Samples","ref_index":2,"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":50,"is_internal_anchor":false},{"citing_arxiv_id":"2605.09927","citing_title":"Information Extraction of Nested Complex Structure of Quantum Cascade Lasers via Large Language Models","ref_index":54,"is_internal_anchor":false},{"citing_arxiv_id":"2604.24902","citing_title":"Safety Drift After Fine-Tuning: Evidence from High-Stakes Domains","ref_index":19,"is_internal_anchor":false}]},"formal_canon":{"evidence_count":0,"sample":[],"anchors":[]},"links":{"html":"https://pith.science/pith/73CMOHP6Z6RAFZDXDV54L7LJKE","json":"https://pith.science/pith/73CMOHP6Z6RAFZDXDV54L7LJKE.json","graph_json":"https://pith.science/api/pith-number/73CMOHP6Z6RAFZDXDV54L7LJKE/graph.json","events_json":"https://pith.science/api/pith-number/73CMOHP6Z6RAFZDXDV54L7LJKE/events.json","paper":"https://pith.science/paper/73CMOHP6"},"agent_actions":{"view_html":"https://pith.science/pith/73CMOHP6Z6RAFZDXDV54L7LJKE","download_json":"https://pith.science/pith/73CMOHP6Z6RAFZDXDV54L7LJKE.json","view_paper":"https://pith.science/paper/73CMOHP6","resolve_alias":"https://pith.science/api/pith-number/resolve?arxiv=2410.15821&json=true","fetch_graph":"https://pith.science/api/pith-number/73CMOHP6Z6RAFZDXDV54L7LJKE/graph.json","fetch_events":"https://pith.science/api/pith-number/73CMOHP6Z6RAFZDXDV54L7LJKE/events.json","actions":{"anchor_timestamp":"https://pith.science/pith/73CMOHP6Z6RAFZDXDV54L7LJKE/action/timestamp_anchor","attest_storage":"https://pith.science/pith/73CMOHP6Z6RAFZDXDV54L7LJKE/action/storage_attestation","attest_author":"https://pith.science/pith/73CMOHP6Z6RAFZDXDV54L7LJKE/action/author_attestation","sign_citation":"https://pith.science/pith/73CMOHP6Z6RAFZDXDV54L7LJKE/action/citation_signature","submit_replication":"https://pith.science/pith/73CMOHP6Z6RAFZDXDV54L7LJKE/action/replication_record"}},"created_at":"2026-07-05T09:23:25.377099+00:00","updated_at":"2026-07-05T09:23:25.377099+00:00"}