{"record_type":"pith_number_record","schema_url":"https://pith.science/schemas/pith-number/v1.json","pith_number":"pith:2024:SP22TY2PMTFJMW42EBIYT3XPPT","short_pith_number":"pith:SP22TY2P","schema_version":"1.0","canonical_sha256":"93f5a9e34f64ca965b9a205189eeef7cc5f170135845c3700170f5e593295d81","source":{"kind":"arxiv","id":"2409.05816","version":2},"attestation_state":"computed","paper":{"title":"Improving Pretraining Data Using Perplexity Correlations","license":"http://creativecommons.org/licenses/by/4.0/","headline":"","cross_cats":["cs.LG","stat.ML"],"primary_cat":"cs.CL","authors_text":"Christopher Potts, Tatsunori Hashimoto, Tristan Thrush","submitted_at":"2024-09-09T17:23:29Z","abstract_excerpt":"Quality pretraining data is often seen as the key to high-performance language models. However, progress in understanding pretraining data has been slow due to the costly pretraining runs required for data selection experiments. We present a framework that avoids these costs and selects high-quality pretraining data without any LLM training of our own. Our work is based on a simple observation: LLM losses on many pretraining texts are correlated with downstream benchmark performance, and selecting high-correlation documents is an effective pretraining data selection method. We build a new stat"},"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":"2409.05816","kind":"arxiv","version":2},"metadata":{"license":"http://creativecommons.org/licenses/by/4.0/","primary_cat":"cs.CL","submitted_at":"2024-09-09T17:23:29Z","cross_cats_sorted":["cs.LG","stat.ML"],"title_canon_sha256":"a675701def5eaaab2ae50b2e3bb3cde31d3bfb23e5902aa8947758c51630e321","abstract_canon_sha256":"286cda8fdc35c9d38dcc2e37f7736b18a12a27f1379897f5e02b953a9dc5cf22"},"schema_version":"1.0"},"receipt":{"kind":"pith_receipt","key_id":"pith-v1-2026-05","algorithm":"ed25519","signed_at":"2026-07-05T10:28:05.018858Z","signature_b64":"hpHC9IwnoaQpJmuycVGhhNh4OUeTSq29ZxMD7n+13dsBnGvdKKdg2FnWn46aEZbHyfjF1MavMfbYlacDlCG1Ag==","signed_message":"canonical_sha256_bytes","builder_version":"pith-number-builder-2026-05-17-v1","receipt_version":"0.3","canonical_sha256":"93f5a9e34f64ca965b9a205189eeef7cc5f170135845c3700170f5e593295d81","last_reissued_at":"2026-07-05T10:28:05.017891Z","signature_status":"signed_v1","first_computed_at":"2026-07-05T10:28:05.017891Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"graph_snapshot":{"paper":{"title":"Improving Pretraining Data Using Perplexity Correlations","license":"http://creativecommons.org/licenses/by/4.0/","headline":"","cross_cats":["cs.LG","stat.ML"],"primary_cat":"cs.CL","authors_text":"Christopher Potts, Tatsunori Hashimoto, Tristan Thrush","submitted_at":"2024-09-09T17:23:29Z","abstract_excerpt":"Quality pretraining data is often seen as the key to high-performance language models. However, progress in understanding pretraining data has been slow due to the costly pretraining runs required for data selection experiments. We present a framework that avoids these costs and selects high-quality pretraining data without any LLM training of our own. Our work is based on a simple observation: LLM losses on many pretraining texts are correlated with downstream benchmark performance, and selecting high-correlation documents is an effective pretraining data selection method. We build a new stat"},"claims":{"count":0,"items":[],"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"source":{"id":"2409.05816","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/2409.05816/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":"2409.05816","created_at":"2026-07-05T10:28:05.018021+00:00"},{"alias_kind":"arxiv_version","alias_value":"2409.05816v2","created_at":"2026-07-05T10:28:05.018021+00:00"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2409.05816","created_at":"2026-07-05T10:28:05.018021+00:00"},{"alias_kind":"pith_short_12","alias_value":"SP22TY2PMTFJ","created_at":"2026-07-05T10:28:05.018021+00:00"},{"alias_kind":"pith_short_16","alias_value":"SP22TY2PMTFJMW42","created_at":"2026-07-05T10:28:05.018021+00:00"},{"alias_kind":"pith_short_8","alias_value":"SP22TY2P","created_at":"2026-07-05T10:28:05.018021+00:00"}],"events":[],"event_summary":{},"paper_claims":[],"inbound_citations":{"count":3,"internal_anchor_count":0,"sample":[{"citing_arxiv_id":"2606.05029","citing_title":"Validity Threats for Foundation Model Research","ref_index":98,"is_internal_anchor":false},{"citing_arxiv_id":"2511.05963","citing_title":"Next-Latent Prediction Transformers Learn Compact World Models","ref_index":29,"is_internal_anchor":false},{"citing_arxiv_id":"2604.11810","citing_title":"GRACE: A Dynamic Coreset Selection Framework for Large Language Model Optimization","ref_index":67,"is_internal_anchor":false}]},"formal_canon":{"evidence_count":0,"sample":[],"anchors":[]},"links":{"html":"https://pith.science/pith/SP22TY2PMTFJMW42EBIYT3XPPT","json":"https://pith.science/pith/SP22TY2PMTFJMW42EBIYT3XPPT.json","graph_json":"https://pith.science/api/pith-number/SP22TY2PMTFJMW42EBIYT3XPPT/graph.json","events_json":"https://pith.science/api/pith-number/SP22TY2PMTFJMW42EBIYT3XPPT/events.json","paper":"https://pith.science/paper/SP22TY2P"},"agent_actions":{"view_html":"https://pith.science/pith/SP22TY2PMTFJMW42EBIYT3XPPT","download_json":"https://pith.science/pith/SP22TY2PMTFJMW42EBIYT3XPPT.json","view_paper":"https://pith.science/paper/SP22TY2P","resolve_alias":"https://pith.science/api/pith-number/resolve?arxiv=2409.05816&json=true","fetch_graph":"https://pith.science/api/pith-number/SP22TY2PMTFJMW42EBIYT3XPPT/graph.json","fetch_events":"https://pith.science/api/pith-number/SP22TY2PMTFJMW42EBIYT3XPPT/events.json","actions":{"anchor_timestamp":"https://pith.science/pith/SP22TY2PMTFJMW42EBIYT3XPPT/action/timestamp_anchor","attest_storage":"https://pith.science/pith/SP22TY2PMTFJMW42EBIYT3XPPT/action/storage_attestation","attest_author":"https://pith.science/pith/SP22TY2PMTFJMW42EBIYT3XPPT/action/author_attestation","sign_citation":"https://pith.science/pith/SP22TY2PMTFJMW42EBIYT3XPPT/action/citation_signature","submit_replication":"https://pith.science/pith/SP22TY2PMTFJMW42EBIYT3XPPT/action/replication_record"}},"created_at":"2026-07-05T10:28:05.018021+00:00","updated_at":"2026-07-05T10:28:05.018021+00:00"}