{"record_type":"pith_number_record","schema_url":"https://pith.science/schemas/pith-number/v1.json","pith_number":"pith:2025:VDYYBT3SBDQBV5VV2JDK3UNEUE","short_pith_number":"pith:VDYYBT3S","schema_version":"1.0","canonical_sha256":"a8f180cf7208e01af6b5d246add1a4a10646fd5e5473cd07eeb087be3ab31c4c","source":{"kind":"arxiv","id":"2503.19206","version":2},"attestation_state":"computed","paper":{"title":"Overtrained Language Models Are Harder to Fine-Tune","license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","headline":"","cross_cats":["cs.AI"],"primary_cat":"cs.CL","authors_text":"Aditi Raghunathan, Graham Neubig, Jacob Mitchell Springer, Kaiyue Wen, Sachin Goyal, Sadhika Malladi, Tanishq Kumar, Xiang Yue","submitted_at":"2025-03-24T23:11:56Z","abstract_excerpt":"Large language models are pre-trained on ever-growing token budgets under the assumption that better pre-training performance translates to improved downstream models. In this work, we challenge this assumption and show that extended pre-training can make models harder to fine-tune, leading to degraded final performance. We term this phenomenon catastrophic overtraining. For example, the instruction-tuned OLMo-1B model pre-trained on 3T tokens leads to over 2% worse performance on multiple standard LLM benchmarks than its 2.3T token counterpart. Through controlled experiments and theoretical a"},"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":"2503.19206","kind":"arxiv","version":2},"metadata":{"license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","primary_cat":"cs.CL","submitted_at":"2025-03-24T23:11:56Z","cross_cats_sorted":["cs.AI"],"title_canon_sha256":"1eca5615d4254a6068f7376be394a53e8f3594a60b4a065e9c79c6e05e7ef40e","abstract_canon_sha256":"5effa2240fbf278bd8a87bd10c3122ba1ff9d301b2d2251742582809469a9839"},"schema_version":"1.0"},"receipt":{"kind":"pith_receipt","key_id":"pith-v1-2026-05","algorithm":"ed25519","signed_at":"2026-07-05T10:40:41.362536Z","signature_b64":"Dn2MuZ2B4SlA2jbngsnjQ2r0ITK9NLT1KVSdA8O/ipk+Eqr/bfvKE6S0NtfIbaKZHNBViOx/e7fLOsmyL4AQBQ==","signed_message":"canonical_sha256_bytes","builder_version":"pith-number-builder-2026-05-17-v1","receipt_version":"0.3","canonical_sha256":"a8f180cf7208e01af6b5d246add1a4a10646fd5e5473cd07eeb087be3ab31c4c","last_reissued_at":"2026-07-05T10:40:41.361873Z","signature_status":"signed_v1","first_computed_at":"2026-07-05T10:40:41.361873Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"graph_snapshot":{"paper":{"title":"Overtrained Language Models Are Harder to Fine-Tune","license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","headline":"","cross_cats":["cs.AI"],"primary_cat":"cs.CL","authors_text":"Aditi Raghunathan, Graham Neubig, Jacob Mitchell Springer, Kaiyue Wen, Sachin Goyal, Sadhika Malladi, Tanishq Kumar, Xiang Yue","submitted_at":"2025-03-24T23:11:56Z","abstract_excerpt":"Large language models are pre-trained on ever-growing token budgets under the assumption that better pre-training performance translates to improved downstream models. In this work, we challenge this assumption and show that extended pre-training can make models harder to fine-tune, leading to degraded final performance. We term this phenomenon catastrophic overtraining. For example, the instruction-tuned OLMo-1B model pre-trained on 3T tokens leads to over 2% worse performance on multiple standard LLM benchmarks than its 2.3T token counterpart. Through controlled experiments and theoretical a"},"claims":{"count":0,"items":[],"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"source":{"id":"2503.19206","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/2503.19206/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":"2503.19206","created_at":"2026-07-05T10:40:41.361954+00:00"},{"alias_kind":"arxiv_version","alias_value":"2503.19206v2","created_at":"2026-07-05T10:40:41.361954+00:00"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2503.19206","created_at":"2026-07-05T10:40:41.361954+00:00"},{"alias_kind":"pith_short_12","alias_value":"VDYYBT3SBDQB","created_at":"2026-07-05T10:40:41.361954+00:00"},{"alias_kind":"pith_short_16","alias_value":"VDYYBT3SBDQBV5VV","created_at":"2026-07-05T10:40:41.361954+00:00"},{"alias_kind":"pith_short_8","alias_value":"VDYYBT3S","created_at":"2026-07-05T10:40:41.361954+00:00"}],"events":[],"event_summary":{},"paper_claims":[],"inbound_citations":{"count":7,"internal_anchor_count":0,"sample":[{"citing_arxiv_id":"2606.25971","citing_title":"Improving Neural Network Training by Decoupling the Magnitude and Direction of Weight Vectors","ref_index":193,"is_internal_anchor":false},{"citing_arxiv_id":"2606.28998","citing_title":"Reward-Free Code Alignment from Pretrained or Fine-Tuned LLM: Unpacking the Trade-offs for Code Generation","ref_index":46,"is_internal_anchor":false},{"citing_arxiv_id":"2605.26097","citing_title":"Forgetting in Language Models: Capacity, Optimization, and Self-Generated Replay","ref_index":55,"is_internal_anchor":false},{"citing_arxiv_id":"2605.23901","citing_title":"LLMs as Noisy Channels: A Shannon Perspective on Model Capacity and Scaling Laws","ref_index":28,"is_internal_anchor":false},{"citing_arxiv_id":"2605.12705","citing_title":"Early Data Exposure Improves Robustness to Subsequent Fine-Tuning","ref_index":19,"is_internal_anchor":false},{"citing_arxiv_id":"2605.06654","citing_title":"Optimizer-Model Consistency: Full Finetuning with the Same Optimizer as Pretraining Forgets Less","ref_index":26,"is_internal_anchor":false},{"citing_arxiv_id":"2604.09258","citing_title":"Nexus: Same Pretraining Loss, Better Downstream Generalization via Common Minima","ref_index":37,"is_internal_anchor":false}]},"formal_canon":{"evidence_count":0,"sample":[],"anchors":[]},"links":{"html":"https://pith.science/pith/VDYYBT3SBDQBV5VV2JDK3UNEUE","json":"https://pith.science/pith/VDYYBT3SBDQBV5VV2JDK3UNEUE.json","graph_json":"https://pith.science/api/pith-number/VDYYBT3SBDQBV5VV2JDK3UNEUE/graph.json","events_json":"https://pith.science/api/pith-number/VDYYBT3SBDQBV5VV2JDK3UNEUE/events.json","paper":"https://pith.science/paper/VDYYBT3S"},"agent_actions":{"view_html":"https://pith.science/pith/VDYYBT3SBDQBV5VV2JDK3UNEUE","download_json":"https://pith.science/pith/VDYYBT3SBDQBV5VV2JDK3UNEUE.json","view_paper":"https://pith.science/paper/VDYYBT3S","resolve_alias":"https://pith.science/api/pith-number/resolve?arxiv=2503.19206&json=true","fetch_graph":"https://pith.science/api/pith-number/VDYYBT3SBDQBV5VV2JDK3UNEUE/graph.json","fetch_events":"https://pith.science/api/pith-number/VDYYBT3SBDQBV5VV2JDK3UNEUE/events.json","actions":{"anchor_timestamp":"https://pith.science/pith/VDYYBT3SBDQBV5VV2JDK3UNEUE/action/timestamp_anchor","attest_storage":"https://pith.science/pith/VDYYBT3SBDQBV5VV2JDK3UNEUE/action/storage_attestation","attest_author":"https://pith.science/pith/VDYYBT3SBDQBV5VV2JDK3UNEUE/action/author_attestation","sign_citation":"https://pith.science/pith/VDYYBT3SBDQBV5VV2JDK3UNEUE/action/citation_signature","submit_replication":"https://pith.science/pith/VDYYBT3SBDQBV5VV2JDK3UNEUE/action/replication_record"}},"created_at":"2026-07-05T10:40:41.361954+00:00","updated_at":"2026-07-05T10:40:41.361954+00:00"}