{"record_type":"pith_number_record","schema_url":"https://pith.science/schemas/pith-number/v1.json","pith_number":"pith:2025:XFDMH72Z24UHOLP6KWQMSNFZED","short_pith_number":"pith:XFDMH72Z","schema_version":"1.0","canonical_sha256":"b946c3ff59d728772dfe55a0c934b920f5171b82373dd5a7a71a210e174fe7fb","source":{"kind":"arxiv","id":"2505.04741","version":1},"attestation_state":"computed","paper":{"title":"When Bad Data Leads to Good Models","license":"http://creativecommons.org/licenses/by/4.0/","headline":"","cross_cats":["cs.AI","cs.CL"],"primary_cat":"cs.LG","authors_text":"Fernanda Vi\\'egas, Kenneth Li, Martin Wattenberg, Yida Chen","submitted_at":"2025-05-07T19:17:49Z","abstract_excerpt":"In large language model (LLM) pretraining, data quality is believed to determine model quality. In this paper, we re-examine the notion of \"quality\" from the perspective of pre- and post-training co-design. Specifically, we explore the possibility that pre-training on more toxic data can lead to better control in post-training, ultimately decreasing a model's output toxicity. First, we use a toy experiment to study how data composition affects the geometry of features in the representation space. Next, through controlled experiments with Olmo-1B models trained on varying ratios of clean and to"},"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":"2505.04741","kind":"arxiv","version":1},"metadata":{"license":"http://creativecommons.org/licenses/by/4.0/","primary_cat":"cs.LG","submitted_at":"2025-05-07T19:17:49Z","cross_cats_sorted":["cs.AI","cs.CL"],"title_canon_sha256":"025b44c66d3ce7023182a2adb700c9a2a28f3b7afd0191d2eb8ff43fef57fb95","abstract_canon_sha256":"a52cf697ac2b208e1db4c45d8f267ccf7e72b17f4c83163abbea522898b2750c"},"schema_version":"1.0"},"receipt":{"kind":"pith_receipt","key_id":"pith-v1-2026-05","algorithm":"ed25519","signed_at":"2026-07-05T11:00:13.005434Z","signature_b64":"6BEK3gGVs02f0bWyLWAyhbdgVaCe8OuPw6e+hHRitejcY8ygsZTDF8FGNY/8BJFFf6nt19wE286oF1J2eDvICw==","signed_message":"canonical_sha256_bytes","builder_version":"pith-number-builder-2026-05-17-v1","receipt_version":"0.3","canonical_sha256":"b946c3ff59d728772dfe55a0c934b920f5171b82373dd5a7a71a210e174fe7fb","last_reissued_at":"2026-07-05T11:00:13.004902Z","signature_status":"signed_v1","first_computed_at":"2026-07-05T11:00:13.004902Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"graph_snapshot":{"paper":{"title":"When Bad Data Leads to Good Models","license":"http://creativecommons.org/licenses/by/4.0/","headline":"","cross_cats":["cs.AI","cs.CL"],"primary_cat":"cs.LG","authors_text":"Fernanda Vi\\'egas, Kenneth Li, Martin Wattenberg, Yida Chen","submitted_at":"2025-05-07T19:17:49Z","abstract_excerpt":"In large language model (LLM) pretraining, data quality is believed to determine model quality. In this paper, we re-examine the notion of \"quality\" from the perspective of pre- and post-training co-design. Specifically, we explore the possibility that pre-training on more toxic data can lead to better control in post-training, ultimately decreasing a model's output toxicity. First, we use a toy experiment to study how data composition affects the geometry of features in the representation space. Next, through controlled experiments with Olmo-1B models trained on varying ratios of clean and to"},"claims":{"count":0,"items":[],"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"source":{"id":"2505.04741","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/2505.04741/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":"2505.04741","created_at":"2026-07-05T11:00:13.004953+00:00"},{"alias_kind":"arxiv_version","alias_value":"2505.04741v1","created_at":"2026-07-05T11:00:13.004953+00:00"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2505.04741","created_at":"2026-07-05T11:00:13.004953+00:00"},{"alias_kind":"pith_short_12","alias_value":"XFDMH72Z24UH","created_at":"2026-07-05T11:00:13.004953+00:00"},{"alias_kind":"pith_short_16","alias_value":"XFDMH72Z24UHOLP6","created_at":"2026-07-05T11:00:13.004953+00:00"},{"alias_kind":"pith_short_8","alias_value":"XFDMH72Z","created_at":"2026-07-05T11:00:13.004953+00:00"}],"events":[],"event_summary":{},"paper_claims":[],"inbound_citations":{"count":1,"internal_anchor_count":0,"sample":[{"citing_arxiv_id":"2606.07387","citing_title":"Making the Most of Limited Data: Score-Aware Training for Text-to-Music Generation","ref_index":5,"is_internal_anchor":false}]},"formal_canon":{"evidence_count":0,"sample":[],"anchors":[]},"links":{"html":"https://pith.science/pith/XFDMH72Z24UHOLP6KWQMSNFZED","json":"https://pith.science/pith/XFDMH72Z24UHOLP6KWQMSNFZED.json","graph_json":"https://pith.science/api/pith-number/XFDMH72Z24UHOLP6KWQMSNFZED/graph.json","events_json":"https://pith.science/api/pith-number/XFDMH72Z24UHOLP6KWQMSNFZED/events.json","paper":"https://pith.science/paper/XFDMH72Z"},"agent_actions":{"view_html":"https://pith.science/pith/XFDMH72Z24UHOLP6KWQMSNFZED","download_json":"https://pith.science/pith/XFDMH72Z24UHOLP6KWQMSNFZED.json","view_paper":"https://pith.science/paper/XFDMH72Z","resolve_alias":"https://pith.science/api/pith-number/resolve?arxiv=2505.04741&json=true","fetch_graph":"https://pith.science/api/pith-number/XFDMH72Z24UHOLP6KWQMSNFZED/graph.json","fetch_events":"https://pith.science/api/pith-number/XFDMH72Z24UHOLP6KWQMSNFZED/events.json","actions":{"anchor_timestamp":"https://pith.science/pith/XFDMH72Z24UHOLP6KWQMSNFZED/action/timestamp_anchor","attest_storage":"https://pith.science/pith/XFDMH72Z24UHOLP6KWQMSNFZED/action/storage_attestation","attest_author":"https://pith.science/pith/XFDMH72Z24UHOLP6KWQMSNFZED/action/author_attestation","sign_citation":"https://pith.science/pith/XFDMH72Z24UHOLP6KWQMSNFZED/action/citation_signature","submit_replication":"https://pith.science/pith/XFDMH72Z24UHOLP6KWQMSNFZED/action/replication_record"}},"created_at":"2026-07-05T11:00:13.004953+00:00","updated_at":"2026-07-05T11:00:13.004953+00:00"}