{"record_type":"pith_number_record","schema_url":"https://pith.science/schemas/pith-number/v1.json","pith_number":"pith:2024:UJXXSEMUFSV3FLT7TK7VRHMFUC","short_pith_number":"pith:UJXXSEMU","schema_version":"1.0","canonical_sha256":"a26f7911942cabb2ae7f9abf589d85a0becc0976eda5dbeede26dec28551f65e","source":{"kind":"arxiv","id":"2407.16607","version":4},"attestation_state":"computed","paper":{"title":"Data Mixture Inference: What do BPE Tokenizers Reveal about their Training Data?","license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","headline":"","cross_cats":["cs.LG"],"primary_cat":"cs.CL","authors_text":"Alisa Liu, Jonathan Hayase, Noah A. Smith, Sewoong Oh, Yejin Choi","submitted_at":"2024-07-23T16:13:22Z","abstract_excerpt":"The pretraining data of today's strongest language models is opaque; in particular, little is known about the proportions of various domains or languages represented. In this work, we tackle a task which we call data mixture inference, which aims to uncover the distributional make-up of training data. We introduce a novel attack based on a previously overlooked source of information: byte-pair encoding (BPE) tokenizers, used by the vast majority of modern language models. Our key insight is that the ordered list of merge rules learned by a BPE tokenizer naturally reveals information about the "},"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":"2407.16607","kind":"arxiv","version":4},"metadata":{"license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","primary_cat":"cs.CL","submitted_at":"2024-07-23T16:13:22Z","cross_cats_sorted":["cs.LG"],"title_canon_sha256":"58f68631091b228389dc12bff3e5788dbd7fc0ba7841838d0f9fc7e6bff20354","abstract_canon_sha256":"335b86f497a45c80eb211d711a640b7ed39d714988b1318ba0d700fed18ed3fa"},"schema_version":"1.0"},"receipt":{"kind":"pith_receipt","key_id":"pith-v1-2026-05","algorithm":"ed25519","signed_at":"2026-07-05T09:42:28.239498Z","signature_b64":"aBZ6aeNSwZ6yT9N57IzlNvux7WCU2ff6/GQUti1t0+GdZkK3O2A3tSihP7RTJOY7Vad2GUkLOtQpTeGROVd1Cg==","signed_message":"canonical_sha256_bytes","builder_version":"pith-number-builder-2026-05-17-v1","receipt_version":"0.3","canonical_sha256":"a26f7911942cabb2ae7f9abf589d85a0becc0976eda5dbeede26dec28551f65e","last_reissued_at":"2026-07-05T09:42:28.239003Z","signature_status":"signed_v1","first_computed_at":"2026-07-05T09:42:28.239003Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"graph_snapshot":{"paper":{"title":"Data Mixture Inference: What do BPE Tokenizers Reveal about their Training Data?","license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","headline":"","cross_cats":["cs.LG"],"primary_cat":"cs.CL","authors_text":"Alisa Liu, Jonathan Hayase, Noah A. Smith, Sewoong Oh, Yejin Choi","submitted_at":"2024-07-23T16:13:22Z","abstract_excerpt":"The pretraining data of today's strongest language models is opaque; in particular, little is known about the proportions of various domains or languages represented. In this work, we tackle a task which we call data mixture inference, which aims to uncover the distributional make-up of training data. We introduce a novel attack based on a previously overlooked source of information: byte-pair encoding (BPE) tokenizers, used by the vast majority of modern language models. Our key insight is that the ordered list of merge rules learned by a BPE tokenizer naturally reveals information about the "},"claims":{"count":0,"items":[],"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"source":{"id":"2407.16607","kind":"arxiv","version":4},"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/2407.16607/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":"2407.16607","created_at":"2026-07-05T09:42:28.239060+00:00"},{"alias_kind":"arxiv_version","alias_value":"2407.16607v4","created_at":"2026-07-05T09:42:28.239060+00:00"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2407.16607","created_at":"2026-07-05T09:42:28.239060+00:00"},{"alias_kind":"pith_short_12","alias_value":"UJXXSEMUFSV3","created_at":"2026-07-05T09:42:28.239060+00:00"},{"alias_kind":"pith_short_16","alias_value":"UJXXSEMUFSV3FLT7","created_at":"2026-07-05T09:42:28.239060+00:00"},{"alias_kind":"pith_short_8","alias_value":"UJXXSEMU","created_at":"2026-07-05T09:42:28.239060+00:00"}],"events":[],"event_summary":{},"paper_claims":[],"inbound_citations":{"count":2,"internal_anchor_count":0,"sample":[{"citing_arxiv_id":"2602.04289","citing_title":"Proxy Compression for Language Modeling","ref_index":7,"is_internal_anchor":false},{"citing_arxiv_id":"2605.13429","citing_title":"TokAlign++: Advancing Vocabulary Adaptation via Better Token Alignment","ref_index":62,"is_internal_anchor":false}]},"formal_canon":{"evidence_count":0,"sample":[],"anchors":[]},"links":{"html":"https://pith.science/pith/UJXXSEMUFSV3FLT7TK7VRHMFUC","json":"https://pith.science/pith/UJXXSEMUFSV3FLT7TK7VRHMFUC.json","graph_json":"https://pith.science/api/pith-number/UJXXSEMUFSV3FLT7TK7VRHMFUC/graph.json","events_json":"https://pith.science/api/pith-number/UJXXSEMUFSV3FLT7TK7VRHMFUC/events.json","paper":"https://pith.science/paper/UJXXSEMU"},"agent_actions":{"view_html":"https://pith.science/pith/UJXXSEMUFSV3FLT7TK7VRHMFUC","download_json":"https://pith.science/pith/UJXXSEMUFSV3FLT7TK7VRHMFUC.json","view_paper":"https://pith.science/paper/UJXXSEMU","resolve_alias":"https://pith.science/api/pith-number/resolve?arxiv=2407.16607&json=true","fetch_graph":"https://pith.science/api/pith-number/UJXXSEMUFSV3FLT7TK7VRHMFUC/graph.json","fetch_events":"https://pith.science/api/pith-number/UJXXSEMUFSV3FLT7TK7VRHMFUC/events.json","actions":{"anchor_timestamp":"https://pith.science/pith/UJXXSEMUFSV3FLT7TK7VRHMFUC/action/timestamp_anchor","attest_storage":"https://pith.science/pith/UJXXSEMUFSV3FLT7TK7VRHMFUC/action/storage_attestation","attest_author":"https://pith.science/pith/UJXXSEMUFSV3FLT7TK7VRHMFUC/action/author_attestation","sign_citation":"https://pith.science/pith/UJXXSEMUFSV3FLT7TK7VRHMFUC/action/citation_signature","submit_replication":"https://pith.science/pith/UJXXSEMUFSV3FLT7TK7VRHMFUC/action/replication_record"}},"created_at":"2026-07-05T09:42:28.239060+00:00","updated_at":"2026-07-05T09:42:28.239060+00:00"}