{"paper":{"title":"Early Data Exposure Improves Robustness to Subsequent Fine-Tuning","license":"http://creativecommons.org/licenses/by/4.0/","headline":"Mixing some target data into pretraining improves retention of that capability after later fine-tuning on new tasks.","cross_cats":[],"primary_cat":"cs.LG","authors_text":"Aditi Raghunathan, Gaurav R. Ghosal, Jacob Mitchell Springer, Lawrence Feng, Ziqian Zhong","submitted_at":"2026-05-12T20:08:00Z","abstract_excerpt":"How can we train models whose post-trained capabilities survive subsequent fine-tuning? Rather than focusing on downstream interventions to mitigate forgetting of upstream capabilities, we study how upstream training choices - that is, the manner in which a capability is acquired - shape how robustly that capability is retained. We investigate this question in a controlled three-stage language-model pipeline: pretraining, post-training to acquire a target capability, and downstream fine-tuning on a new objective. Across 135M and 1B models, two post-training domains, and two downstream fine-tun"},"claims":{"count":4,"items":[{"kind":"strongest_claim","text":"early exposure - mixing post-training data into pretraining - consistently improves the frontier between retained upstream performance and downstream performance. In compute-matched experiments, where the target data must be allocated between pretraining and post-training, we find that the optimum lies at neither extreme.","source":"verdict.strongest_claim","status":"machine_extracted","claim_id":"C1","attestation":"unclaimed"},{"kind":"weakest_assumption","text":"That the controlled three-stage pipeline and specific model sizes/tasks used here capture the dynamics that matter in larger-scale, real-world training where data distributions and objectives are more complex and less cleanly separated.","source":"verdict.weakest_assumption","status":"machine_extracted","claim_id":"C2","attestation":"unclaimed"},{"kind":"one_line_summary","text":"Early mixing of post-training data into pretraining improves retention of acquired capabilities after subsequent fine-tuning in language models.","source":"verdict.one_line_summary","status":"machine_extracted","claim_id":"C3","attestation":"unclaimed"},{"kind":"headline","text":"Mixing some target data into pretraining improves retention of that capability after later fine-tuning on new tasks.","source":"verdict.pith_extraction.headline","status":"machine_extracted","claim_id":"C4","attestation":"unclaimed"}],"snapshot_sha256":"49e4a956cd8d03ee3867551d08e960145732db5c04d665042db2fec67d66b0cf"},"source":{"id":"2605.12705","kind":"arxiv","version":1},"verdict":{"id":"5e444599-3348-46fb-8ccf-109ac0cf67b1","model_set":{"reader":"grok-4.3"},"created_at":"2026-05-14T20:40:28.418904Z","strongest_claim":"early exposure - mixing post-training data into pretraining - consistently improves the frontier between retained upstream performance and downstream performance. In compute-matched experiments, where the target data must be allocated between pretraining and post-training, we find that the optimum lies at neither extreme.","one_line_summary":"Early mixing of post-training data into pretraining improves retention of acquired capabilities after subsequent fine-tuning in language models.","pipeline_version":"pith-pipeline@v0.9.0","weakest_assumption":"That the controlled three-stage pipeline and specific model sizes/tasks used here capture the dynamics that matter in larger-scale, real-world training where data distributions and objectives are more complex and less cleanly separated.","pith_extraction_headline":"Mixing some target data into pretraining improves retention of that capability after later fine-tuning on new tasks."},"references":{"count":25,"sample":[{"doi":"","year":null,"title":"SmolLM2: When Smol Goes Big -- Data-Centric Training of a Small Language Model","work_id":"8472f581-14d4-40f8-8189-62ed8b470c4e","ref_index":1,"cited_arxiv_id":"2502.02737","is_internal_anchor":true},{"doi":"","year":null,"title":"Louis Bethune, David Grangier, Dan Busbridge, Eleonora Gualdoni, Marco Cuturi, and Pierre Ablin","work_id":"253db970-0a8b-4c85-85b9-b762b6ce3086","ref_index":2,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":null,"title":"arXiv preprint arXiv:2502.06042 , year=","work_id":"9617c4ed-1474-439a-9d08-e0612fa954ae","ref_index":3,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":null,"title":"Lora learns less and forgets less","work_id":"ca00d37c-5297-4fa7-b692-52260eb614f2","ref_index":4,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":null,"title":"Scaling laws for predicting downstream performance in llms, 2025","work_id":"10184b86-cdee-4047-93c1-53dd0f8c9a18","ref_index":5,"cited_arxiv_id":"","is_internal_anchor":false}],"resolved_work":25,"snapshot_sha256":"03edf1716805044acdfc9f8e36e50bd3febc707645515dfd849d5d2aa47dfdc4","internal_anchors":7},"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"}