{"paper":{"title":"Mix, Don't Tune: Bilingual Pre-Training Outperforms Hyperparameter Search in Data-Constrained Settings","license":"http://creativecommons.org/licenses/by/4.0/","headline":"Mixing high-resource language data outperforms hyperparameter tuning for low-resource pre-training.","cross_cats":[],"primary_cat":"cs.LG","authors_text":"Anastasiia Sedova, Jes Frellsen, Louis B\\'ethune, Natalie Schluter, Paul Jeha, Pierre Ablin, Skyler Seto","submitted_at":"2026-05-13T09:17:51Z","abstract_excerpt":"For most languages of the world, language model pre-training operates in a data-constrained regime where models must repeat their training data many times, degrading generalization. Two remedies exist: aggressive hyperparameter tuning such as high weight decay, and mixing in data from a high-resource auxiliary language to directly aid the low-resource target. While hyperparameter tuning regularizes the model by shrinking weights to restrict network capacity, auxiliary data mixing uses a tunable mixing ratio to expand the training distribution and diversify the training signal with new knowledg"},"claims":{"count":4,"items":[{"kind":"strongest_claim","text":"mixing yields larger improvements than hyperparameter tuning on both validation loss and downstream task accuracy, and the gap grows with model size. We quantify how much mixing helps: it boosts performance by an amount equivalent to 2--3× the unique target data on validation loss and 2--13× on downstream task accuracy, with the gain scaling steeply with model size.","source":"verdict.strongest_claim","status":"machine_extracted","claim_id":"C1","attestation":"unclaimed"},{"kind":"weakest_assumption","text":"That the chosen mixing ratios are near-optimal and that English data supplies useful, non-conflicting signal for Arabic without introducing domain mismatch that would require separate controls.","source":"verdict.weakest_assumption","status":"machine_extracted","claim_id":"C2","attestation":"unclaimed"},{"kind":"one_line_summary","text":"Mixing auxiliary high-resource language data outperforms hyperparameter tuning in data-constrained bilingual pre-training, with gains equivalent to 2-13 times more unique target data.","source":"verdict.one_line_summary","status":"machine_extracted","claim_id":"C3","attestation":"unclaimed"},{"kind":"headline","text":"Mixing high-resource language data outperforms hyperparameter tuning for low-resource pre-training.","source":"verdict.pith_extraction.headline","status":"machine_extracted","claim_id":"C4","attestation":"unclaimed"}],"snapshot_sha256":"e96f95120f13970c772a7c89914519d6f6af522e48cbbb2e517079cc756cfe12"},"source":{"id":"2605.13225","kind":"arxiv","version":1},"verdict":{"id":"a388c296-c723-4ebd-9e20-b9173d344f72","model_set":{"reader":"grok-4.3"},"created_at":"2026-05-14T20:13:37.436444Z","strongest_claim":"mixing yields larger improvements than hyperparameter tuning on both validation loss and downstream task accuracy, and the gap grows with model size. We quantify how much mixing helps: it boosts performance by an amount equivalent to 2--3× the unique target data on validation loss and 2--13× on downstream task accuracy, with the gain scaling steeply with model size.","one_line_summary":"Mixing auxiliary high-resource language data outperforms hyperparameter tuning in data-constrained bilingual pre-training, with gains equivalent to 2-13 times more unique target data.","pipeline_version":"pith-pipeline@v0.9.0","weakest_assumption":"That the chosen mixing ratios are near-optimal and that English data supplies useful, non-conflicting signal for Arabic without introducing domain mismatch that would require separate controls.","pith_extraction_headline":"Mixing high-resource language data outperforms hyperparameter tuning for low-resource pre-training."},"references":{"count":32,"sample":[{"doi":"","year":null,"title":"Peter Clark, Isaac Cowhey, Oren Etzioni, Tushar Khot, Ashish Sabharwal, Carissa Schoenick, and Oyvind Tafjord","work_id":"b14fac55-a32e-45fb-b906-93ca894c02d8","ref_index":1,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":1911,"title":"Unsupervised Cross-lingual Representation Learning at Scale","work_id":"32df83f5-69ea-418b-9a8d-03c2d6695b80","ref_index":2,"cited_arxiv_id":"1911.02116","is_internal_anchor":true},{"doi":"","year":null,"title":"arXiv preprint arXiv:2310.05492 , year=","work_id":"cc28e604-660c-4b03-853d-dec1e3dac25f","ref_index":3,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":null,"title":"arXiv preprint arXiv:2403.08540 (2024)","work_id":"b3ccca34-2e12-48b4-ad09-521ec9797b0c","ref_index":4,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":null,"title":"Dan Hendrycks, Collin Burns, Steven Basart, Andy Zou, Mantas Mazeika, Dawn Song, and Jacob Steinhardt","work_id":"17ddc4a7-0ce2-4768-a1e4-b5d692e498e2","ref_index":5,"cited_arxiv_id":"","is_internal_anchor":false}],"resolved_work":32,"snapshot_sha256":"7ce333095a4a04d9987e66ed9e4c0e50751a278062957162ba4f7ccf4b0050c9","internal_anchors":10},"formal_canon":{"evidence_count":2,"snapshot_sha256":"10c417e86119014ae421d4f9069f9f5c0389799516d02066f38d0fb255324a0c"},"author_claims":{"count":0,"strong_count":0,"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"builder_version":"pith-number-builder-2026-05-17-v1"}