{"record_type":"pith_number_record","schema_url":"https://pith.science/schemas/pith-number/v1.json","pith_number":"pith:2022:YNRPOVXKQDJHZE2AMXWQQNQL6E","short_pith_number":"pith:YNRPOVXK","schema_version":"1.0","canonical_sha256":"c362f756ea80d27c934065ed08360bf11eb698975ccfb934d0523dc42e361a3e","source":{"kind":"arxiv","id":"2203.05482","version":3},"attestation_state":"computed","paper":{"title":"Model soups: averaging weights of multiple fine-tuned models improves accuracy without increasing inference time","license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","headline":"","cross_cats":["cs.CL","cs.CV"],"primary_cat":"cs.LG","authors_text":"Ali Farhadi, Ari S. Morcos, Gabriel Ilharco, Hongseok Namkoong, Ludwig Schmidt, Mitchell Wortsman, Raphael Gontijo-Lopes, Rebecca Roelofs, Samir Yitzhak Gadre, Simon Kornblith, Yair Carmon","submitted_at":"2022-03-10T17:03:49Z","abstract_excerpt":"The conventional recipe for maximizing model accuracy is to (1) train multiple models with various hyperparameters and (2) pick the individual model which performs best on a held-out validation set, discarding the remainder. In this paper, we revisit the second step of this procedure in the context of fine-tuning large pre-trained models, where fine-tuned models often appear to lie in a single low error basin. We show that averaging the weights of multiple models fine-tuned with different hyperparameter configurations often improves accuracy and robustness. Unlike a conventional ensemble, we m"},"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":"2203.05482","kind":"arxiv","version":3},"metadata":{"license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","primary_cat":"cs.LG","submitted_at":"2022-03-10T17:03:49Z","cross_cats_sorted":["cs.CL","cs.CV"],"title_canon_sha256":"51862f7d7f6becfeb6cdc9486965f7d936e4c3108bcb6fa4c70f9b29473872a0","abstract_canon_sha256":"ac88a086ddf3e7eda025b9c0225df91ca0fbcf234595736ffb5a82ec4ab2f60f"},"schema_version":"1.0"},"receipt":{"kind":"pith_receipt","key_id":"pith-v1-2026-05","algorithm":"ed25519","signed_at":"2026-07-05T04:36:47.860381Z","signature_b64":"udvpYZqmP54niAV2M0q1gqaNMbDfp0qT+wWgKL4YRo7errxpa2p7bRROBDrN6CIYaEIJQo3T5X6HXwF/mBA8BA==","signed_message":"canonical_sha256_bytes","builder_version":"pith-number-builder-2026-05-17-v1","receipt_version":"0.3","canonical_sha256":"c362f756ea80d27c934065ed08360bf11eb698975ccfb934d0523dc42e361a3e","last_reissued_at":"2026-07-05T04:36:47.859955Z","signature_status":"signed_v1","first_computed_at":"2026-07-05T04:36:47.859955Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"graph_snapshot":{"paper":{"title":"Model soups: averaging weights of multiple fine-tuned models improves accuracy without increasing inference time","license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","headline":"","cross_cats":["cs.CL","cs.CV"],"primary_cat":"cs.LG","authors_text":"Ali Farhadi, Ari S. Morcos, Gabriel Ilharco, Hongseok Namkoong, Ludwig Schmidt, Mitchell Wortsman, Raphael Gontijo-Lopes, Rebecca Roelofs, Samir Yitzhak Gadre, Simon Kornblith, Yair Carmon","submitted_at":"2022-03-10T17:03:49Z","abstract_excerpt":"The conventional recipe for maximizing model accuracy is to (1) train multiple models with various hyperparameters and (2) pick the individual model which performs best on a held-out validation set, discarding the remainder. In this paper, we revisit the second step of this procedure in the context of fine-tuning large pre-trained models, where fine-tuned models often appear to lie in a single low error basin. We show that averaging the weights of multiple models fine-tuned with different hyperparameter configurations often improves accuracy and robustness. Unlike a conventional ensemble, we m"},"claims":{"count":0,"items":[],"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"source":{"id":"2203.05482","kind":"arxiv","version":3},"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/2203.05482/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":"2203.05482","created_at":"2026-07-05T04:36:47.860011+00:00"},{"alias_kind":"arxiv_version","alias_value":"2203.05482v3","created_at":"2026-07-05T04:36:47.860011+00:00"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2203.05482","created_at":"2026-07-05T04:36:47.860011+00:00"},{"alias_kind":"pith_short_12","alias_value":"YNRPOVXKQDJH","created_at":"2026-07-05T04:36:47.860011+00:00"},{"alias_kind":"pith_short_16","alias_value":"YNRPOVXKQDJHZE2A","created_at":"2026-07-05T04:36:47.860011+00:00"},{"alias_kind":"pith_short_8","alias_value":"YNRPOVXK","created_at":"2026-07-05T04:36:47.860011+00:00"}],"events":[],"event_summary":{},"paper_claims":[],"inbound_citations":{"count":12,"internal_anchor_count":0,"sample":[{"citing_arxiv_id":"2606.10929","citing_title":"Recoverable but Not Stationary:Local Linear Structures in Weights and Activations","ref_index":20,"is_internal_anchor":false},{"citing_arxiv_id":"2606.05016","citing_title":"TaDA: Calibrated Probe Gating for Task-Domain LoRA Merging","ref_index":21,"is_internal_anchor":false},{"citing_arxiv_id":"2402.05070","citing_title":"A Roadmap to Pluralistic Alignment","ref_index":295,"is_internal_anchor":false},{"citing_arxiv_id":"2509.20354","citing_title":"EmbeddingGemma: Powerful and Lightweight Text Representations","ref_index":25,"is_internal_anchor":false},{"citing_arxiv_id":"2205.01917","citing_title":"CoCa: Contrastive Captioners are Image-Text Foundation Models","ref_index":52,"is_internal_anchor":false},{"citing_arxiv_id":"2605.12705","citing_title":"Early Data Exposure Improves Robustness to Subsequent Fine-Tuning","ref_index":20,"is_internal_anchor":false},{"citing_arxiv_id":"2212.04089","citing_title":"Editing Models with Task Arithmetic","ref_index":106,"is_internal_anchor":false},{"citing_arxiv_id":"2604.27308","citing_title":"BoostLoRA: Growing Effective Rank by Boosting Adapters","ref_index":34,"is_internal_anchor":false},{"citing_arxiv_id":"2604.26689","citing_title":"Atomic-Probe Governance for Skill Updates in Compositional Robot Policies","ref_index":48,"is_internal_anchor":false},{"citing_arxiv_id":"2204.14198","citing_title":"Flamingo: a Visual Language Model for Few-Shot Learning","ref_index":128,"is_internal_anchor":false},{"citing_arxiv_id":"2604.26689","citing_title":"Atomic-Probe Governance for Skill Updates in Compositional Robot Policies","ref_index":48,"is_internal_anchor":false},{"citing_arxiv_id":"2604.18473","citing_title":"Train Separately, Merge Together: Modular Post-Training with Mixture-of-Experts","ref_index":44,"is_internal_anchor":false}]},"formal_canon":{"evidence_count":0,"sample":[],"anchors":[]},"links":{"html":"https://pith.science/pith/YNRPOVXKQDJHZE2AMXWQQNQL6E","json":"https://pith.science/pith/YNRPOVXKQDJHZE2AMXWQQNQL6E.json","graph_json":"https://pith.science/api/pith-number/YNRPOVXKQDJHZE2AMXWQQNQL6E/graph.json","events_json":"https://pith.science/api/pith-number/YNRPOVXKQDJHZE2AMXWQQNQL6E/events.json","paper":"https://pith.science/paper/YNRPOVXK"},"agent_actions":{"view_html":"https://pith.science/pith/YNRPOVXKQDJHZE2AMXWQQNQL6E","download_json":"https://pith.science/pith/YNRPOVXKQDJHZE2AMXWQQNQL6E.json","view_paper":"https://pith.science/paper/YNRPOVXK","resolve_alias":"https://pith.science/api/pith-number/resolve?arxiv=2203.05482&json=true","fetch_graph":"https://pith.science/api/pith-number/YNRPOVXKQDJHZE2AMXWQQNQL6E/graph.json","fetch_events":"https://pith.science/api/pith-number/YNRPOVXKQDJHZE2AMXWQQNQL6E/events.json","actions":{"anchor_timestamp":"https://pith.science/pith/YNRPOVXKQDJHZE2AMXWQQNQL6E/action/timestamp_anchor","attest_storage":"https://pith.science/pith/YNRPOVXKQDJHZE2AMXWQQNQL6E/action/storage_attestation","attest_author":"https://pith.science/pith/YNRPOVXKQDJHZE2AMXWQQNQL6E/action/author_attestation","sign_citation":"https://pith.science/pith/YNRPOVXKQDJHZE2AMXWQQNQL6E/action/citation_signature","submit_replication":"https://pith.science/pith/YNRPOVXKQDJHZE2AMXWQQNQL6E/action/replication_record"}},"created_at":"2026-07-05T04:36:47.860011+00:00","updated_at":"2026-07-05T04:36:47.860011+00:00"}