{"record_type":"pith_number_record","schema_url":"https://pith.science/schemas/pith-number/v1.json","pith_number":"pith:2026:2T5IQMSX4MYJJU2DXQ4W55YZFD","short_pith_number":"pith:2T5IQMSX","schema_version":"1.0","canonical_sha256":"d4fa883257e33094d343bc396ef71928d8522a248c903bea260f58104eb70458","source":{"kind":"arxiv","id":"2605.03226","version":2},"attestation_state":"computed","paper":{"title":"Self-Mined Hardness for Safety Fine-Tuning","license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","headline":"Models can improve safety fine-tuning by selecting prompts based on how often their own responses are judged harmful.","cross_cats":["cs.AI","cs.CR"],"primary_cat":"cs.LG","authors_text":"Donghua Zhang, Garv Shah, Prakhar Gupta","submitted_at":"2026-05-04T23:30:29Z","abstract_excerpt":"Safety fine-tuning of language models typically requires a curated adversarial dataset. We take a different approach: score each candidate prompt's difficulty by how often the target model's own rollouts are judged harmful, then fine-tune on the hardest prompts paired with the model's own non-jailbroken rollouts. On Llama-3-8B-Instruct and Llama-3.2-3B-Instruct, this approach cuts the WildJailbreak attack success rate from 11.5% and 20.1% down to 1-3%, but pushes refusal on jailbreak-shaped benign prompts from 14-22% to 74-94%. Interleaving the same hard prompts 1:1 with adversarially-framed b"},"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":true},"canonical_record":{"source":{"id":"2605.03226","kind":"arxiv","version":2},"metadata":{"license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","primary_cat":"cs.LG","submitted_at":"2026-05-04T23:30:29Z","cross_cats_sorted":["cs.AI","cs.CR"],"title_canon_sha256":"8960708f235e99c37ce82d273ec3d0dfc3d5230cb674c314ecd322d93ff6ea0f","abstract_canon_sha256":"2a3f685c21e3536239d4e7eb3c3674fabea9b16ff3ba2c2182c8d5a13c2d0909"},"schema_version":"1.0"},"receipt":{"kind":"pith_receipt","key_id":"pith-v1-2026-05","algorithm":"ed25519","signed_at":"2026-06-09T02:08:43.560701Z","signature_b64":"OlfKrvfBpTKdWwQ/ZbHvYWPUF/qRiDhUqfIq9wH+7IiIqsTH0nUvAqlnvxDzW+gXVkV38oYNT+matTbr3+nrBg==","signed_message":"canonical_sha256_bytes","builder_version":"pith-number-builder-2026-05-17-v1","receipt_version":"0.3","canonical_sha256":"d4fa883257e33094d343bc396ef71928d8522a248c903bea260f58104eb70458","last_reissued_at":"2026-06-09T02:08:43.559849Z","signature_status":"signed_v1","first_computed_at":"2026-06-09T02:08:43.559849Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"graph_snapshot":{"paper":{"title":"Self-Mined Hardness for Safety Fine-Tuning","license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","headline":"Models can improve safety fine-tuning by selecting prompts based on how often their own responses are judged harmful.","cross_cats":["cs.AI","cs.CR"],"primary_cat":"cs.LG","authors_text":"Donghua Zhang, Garv Shah, Prakhar Gupta","submitted_at":"2026-05-04T23:30:29Z","abstract_excerpt":"Safety fine-tuning of language models typically requires a curated adversarial dataset. We take a different approach: score each candidate prompt's difficulty by how often the target model's own rollouts are judged harmful, then fine-tune on the hardest prompts paired with the model's own non-jailbroken rollouts. On Llama-3-8B-Instruct and Llama-3.2-3B-Instruct, this approach cuts the WildJailbreak attack success rate from 11.5% and 20.1% down to 1-3%, but pushes refusal on jailbreak-shaped benign prompts from 14-22% to 74-94%. Interleaving the same hard prompts 1:1 with adversarially-framed b"},"claims":{"count":4,"items":[{"kind":"strongest_claim","text":"On Llama-3-8B-Instruct and Llama-3.2-3B-Instruct, this approach cuts the WildJailbreak attack success rate from 11.5% and 20.1% down to 1-3%, but pushes refusal on jailbreak-shaped benign prompts from 14-22% to 74-94%. Interleaving the same hard prompts 1:1 with adversarially-framed benign prompts cuts that refusal back down to 30-51% on 8B and 52-72% on 3B, at a cost of 2-6 percentage points of attack success rate.","source":"verdict.strongest_claim","status":"machine_extracted","claim_id":"C1","attestation":"unclaimed"},{"kind":"weakest_assumption","text":"That the frequency with which the model's own rollouts are judged harmful provides a reliable and unbiased measure of prompt difficulty for safety fine-tuning, and that the external harm judgment itself is consistent and free of systematic error.","source":"verdict.weakest_assumption","status":"machine_extracted","claim_id":"C2","attestation":"unclaimed"},{"kind":"one_line_summary","text":"Self-mined hardness from model rollouts reduces WildJailbreak attack success rates to 1-3% on Llama models but increases over-refusal on benign prompts, which mixing with adversarially-framed benign prompts partially mitigates.","source":"verdict.one_line_summary","status":"machine_extracted","claim_id":"C3","attestation":"unclaimed"},{"kind":"headline","text":"Models can improve safety fine-tuning by selecting prompts based on how often their own responses are judged harmful.","source":"verdict.pith_extraction.headline","status":"machine_extracted","claim_id":"C4","attestation":"unclaimed"}],"snapshot_sha256":"3d7e31b15bb10f819dab1db20e5f3ebbcba4f29c74aeb78ea690417b533169d4"},"source":{"id":"2605.03226","kind":"arxiv","version":2},"verdict":{"id":"8e991f83-c42a-4de8-8aef-93c71675d900","model_set":{"reader":"grok-4.3"},"created_at":"2026-05-08T18:17:53.452364Z","strongest_claim":"On Llama-3-8B-Instruct and Llama-3.2-3B-Instruct, this approach cuts the WildJailbreak attack success rate from 11.5% and 20.1% down to 1-3%, but pushes refusal on jailbreak-shaped benign prompts from 14-22% to 74-94%. Interleaving the same hard prompts 1:1 with adversarially-framed benign prompts cuts that refusal back down to 30-51% on 8B and 52-72% on 3B, at a cost of 2-6 percentage points of attack success rate.","one_line_summary":"Self-mined hardness from model rollouts reduces WildJailbreak attack success rates to 1-3% on Llama models but increases over-refusal on benign prompts, which mixing with adversarially-framed benign prompts partially mitigates.","pipeline_version":"pith-pipeline@v0.9.0","weakest_assumption":"That the frequency with which the model's own rollouts are judged harmful provides a reliable and unbiased measure of prompt difficulty for safety fine-tuning, and that the external harm judgment itself is consistent and free of systematic error.","pith_extraction_headline":"Models can improve safety fine-tuning by selecting prompts based on how often their own responses are judged harmful."},"integrity":{"clean":true,"summary":{"advisory":0,"critical":0,"by_detector":{},"informational":0},"endpoint":"/pith/2605.03226/integrity.json","findings":[],"available":true,"detectors_run":[{"name":"ai_meta_artifact","ran_at":"2026-05-20T14:35:20.977131Z","status":"completed","version":"1.0.0","findings_count":0},{"name":"doi_title_agreement","ran_at":"2026-05-20T01:31:21.804826Z","status":"completed","version":"1.0.0","findings_count":0},{"name":"doi_compliance","ran_at":"2026-05-19T15:34:04.492328Z","status":"completed","version":"1.0.0","findings_count":0}],"snapshot_sha256":"63b72441310a29bbde61c85889cfa895137a0ec66dbff06e0cbab195626790f7"},"references":{"count":0,"sample":[],"resolved_work":0,"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57","internal_anchors":0},"formal_canon":{"evidence_count":2,"snapshot_sha256":"cff818b5e7fc17a1ca754d37c5d18dd7efd6104dc5d1512a4b9080fdd2460396"},"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":"2605.03226","created_at":"2026-06-09T02:08:43.559980+00:00"},{"alias_kind":"arxiv_version","alias_value":"2605.03226v2","created_at":"2026-06-09T02:08:43.559980+00:00"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2605.03226","created_at":"2026-06-09T02:08:43.559980+00:00"},{"alias_kind":"pith_short_12","alias_value":"2T5IQMSX4MYJ","created_at":"2026-06-09T02:08:43.559980+00:00"},{"alias_kind":"pith_short_16","alias_value":"2T5IQMSX4MYJJU2D","created_at":"2026-06-09T02:08:43.559980+00:00"},{"alias_kind":"pith_short_8","alias_value":"2T5IQMSX","created_at":"2026-06-09T02:08:43.559980+00:00"}],"events":[],"event_summary":{},"paper_claims":[],"inbound_citations":{"count":0,"internal_anchor_count":0,"sample":[]},"formal_canon":{"evidence_count":2,"sample":[],"anchors":[]},"links":{"html":"https://pith.science/pith/2T5IQMSX4MYJJU2DXQ4W55YZFD","json":"https://pith.science/pith/2T5IQMSX4MYJJU2DXQ4W55YZFD.json","graph_json":"https://pith.science/api/pith-number/2T5IQMSX4MYJJU2DXQ4W55YZFD/graph.json","events_json":"https://pith.science/api/pith-number/2T5IQMSX4MYJJU2DXQ4W55YZFD/events.json","paper":"https://pith.science/paper/2T5IQMSX"},"agent_actions":{"view_html":"https://pith.science/pith/2T5IQMSX4MYJJU2DXQ4W55YZFD","download_json":"https://pith.science/pith/2T5IQMSX4MYJJU2DXQ4W55YZFD.json","view_paper":"https://pith.science/paper/2T5IQMSX","resolve_alias":"https://pith.science/api/pith-number/resolve?arxiv=2605.03226&json=true","fetch_graph":"https://pith.science/api/pith-number/2T5IQMSX4MYJJU2DXQ4W55YZFD/graph.json","fetch_events":"https://pith.science/api/pith-number/2T5IQMSX4MYJJU2DXQ4W55YZFD/events.json","actions":{"anchor_timestamp":"https://pith.science/pith/2T5IQMSX4MYJJU2DXQ4W55YZFD/action/timestamp_anchor","attest_storage":"https://pith.science/pith/2T5IQMSX4MYJJU2DXQ4W55YZFD/action/storage_attestation","attest_author":"https://pith.science/pith/2T5IQMSX4MYJJU2DXQ4W55YZFD/action/author_attestation","sign_citation":"https://pith.science/pith/2T5IQMSX4MYJJU2DXQ4W55YZFD/action/citation_signature","submit_replication":"https://pith.science/pith/2T5IQMSX4MYJJU2DXQ4W55YZFD/action/replication_record"}},"created_at":"2026-06-09T02:08:43.559980+00:00","updated_at":"2026-06-09T02:08:43.559980+00:00"}