{"record_type":"pith_number_record","schema_url":"https://pith.science/schemas/pith-number/v1.json","pith_number":"pith:2022:LS6BLTX75J6WZBNC7W4MRMGNYZ","short_pith_number":"pith:LS6BLTX7","schema_version":"1.0","canonical_sha256":"5cbc15ceffea7d6c85a2fdb8c8b0cdc6739f9121cb13e199588bead175db9dd6","source":{"kind":"arxiv","id":"2209.14981","version":2},"attestation_state":"computed","paper":{"title":"Stop Wasting My Time! Saving Days of ImageNet and BERT Training with Latest Weight Averaging","license":"http://creativecommons.org/licenses/by/4.0/","headline":"","cross_cats":["cs.AI","stat.ML"],"primary_cat":"cs.LG","authors_text":"Jean Kaddour","submitted_at":"2022-09-29T17:46:13Z","abstract_excerpt":"Training vision or language models on large datasets can take days, if not weeks. We show that averaging the weights of the k latest checkpoints, each collected at the end of an epoch, can speed up the training progression in terms of loss and accuracy by dozens of epochs, corresponding to time savings up to ~68 and ~30 GPU hours when training a ResNet50 on ImageNet and RoBERTa-Base model on WikiText-103, respectively. We also provide the code and model checkpoint trajectory to reproduce the results and facilitate research on reusing historical weights for faster convergence."},"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":"2209.14981","kind":"arxiv","version":2},"metadata":{"license":"http://creativecommons.org/licenses/by/4.0/","primary_cat":"cs.LG","submitted_at":"2022-09-29T17:46:13Z","cross_cats_sorted":["cs.AI","stat.ML"],"title_canon_sha256":"99ed153552f500fcda56e0b8f3340b3d97fd5520a78d37d836266b2f2d9d55fd","abstract_canon_sha256":"f9e74f3a51912bf142df159c2f97186c960b0e0342cfd7c15d11d8ea6d0a9666"},"schema_version":"1.0"},"receipt":{"kind":"pith_receipt","key_id":"pith-v1-2026-05","algorithm":"ed25519","signed_at":"2026-07-05T05:04:04.076283Z","signature_b64":"o0Td/bH0BgyGW07bVTSZDOV7Z28AmZh9Dzx7OtmL5CgSd8C2eE4vLszAwZxafLyNgh7o9zmPCsHOPo8WRs0yDA==","signed_message":"canonical_sha256_bytes","builder_version":"pith-number-builder-2026-05-17-v1","receipt_version":"0.3","canonical_sha256":"5cbc15ceffea7d6c85a2fdb8c8b0cdc6739f9121cb13e199588bead175db9dd6","last_reissued_at":"2026-07-05T05:04:04.075862Z","signature_status":"signed_v1","first_computed_at":"2026-07-05T05:04:04.075862Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"graph_snapshot":{"paper":{"title":"Stop Wasting My Time! Saving Days of ImageNet and BERT Training with Latest Weight Averaging","license":"http://creativecommons.org/licenses/by/4.0/","headline":"","cross_cats":["cs.AI","stat.ML"],"primary_cat":"cs.LG","authors_text":"Jean Kaddour","submitted_at":"2022-09-29T17:46:13Z","abstract_excerpt":"Training vision or language models on large datasets can take days, if not weeks. We show that averaging the weights of the k latest checkpoints, each collected at the end of an epoch, can speed up the training progression in terms of loss and accuracy by dozens of epochs, corresponding to time savings up to ~68 and ~30 GPU hours when training a ResNet50 on ImageNet and RoBERTa-Base model on WikiText-103, respectively. We also provide the code and model checkpoint trajectory to reproduce the results and facilitate research on reusing historical weights for faster convergence."},"claims":{"count":0,"items":[],"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"source":{"id":"2209.14981","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/2209.14981/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":"2209.14981","created_at":"2026-07-05T05:04:04.075922+00:00"},{"alias_kind":"arxiv_version","alias_value":"2209.14981v2","created_at":"2026-07-05T05:04:04.075922+00:00"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2209.14981","created_at":"2026-07-05T05:04:04.075922+00:00"},{"alias_kind":"pith_short_12","alias_value":"LS6BLTX75J6W","created_at":"2026-07-05T05:04:04.075922+00:00"},{"alias_kind":"pith_short_16","alias_value":"LS6BLTX75J6WZBNC","created_at":"2026-07-05T05:04:04.075922+00:00"},{"alias_kind":"pith_short_8","alias_value":"LS6BLTX7","created_at":"2026-07-05T05:04:04.075922+00:00"}],"events":[],"event_summary":{},"paper_claims":[],"inbound_citations":{"count":4,"internal_anchor_count":0,"sample":[{"citing_arxiv_id":"2606.25086","citing_title":"Training for the Model You Return: Improving Optimization for Iterate-Averaged Language Models","ref_index":17,"is_internal_anchor":false},{"citing_arxiv_id":"2606.25086","citing_title":"Training for the Model You Return: Improving Optimization for Iterate-Averaged Language Models","ref_index":17,"is_internal_anchor":false},{"citing_arxiv_id":"2604.02719","citing_title":"MOMO: Mars Orbital Model Foundation Model for Mars Orbital Applications","ref_index":35,"is_internal_anchor":false},{"citing_arxiv_id":"2502.05171","citing_title":"Scaling up Test-Time Compute with Latent Reasoning: A Recurrent Depth Approach","ref_index":78,"is_internal_anchor":false}]},"formal_canon":{"evidence_count":0,"sample":[],"anchors":[]},"links":{"html":"https://pith.science/pith/LS6BLTX75J6WZBNC7W4MRMGNYZ","json":"https://pith.science/pith/LS6BLTX75J6WZBNC7W4MRMGNYZ.json","graph_json":"https://pith.science/api/pith-number/LS6BLTX75J6WZBNC7W4MRMGNYZ/graph.json","events_json":"https://pith.science/api/pith-number/LS6BLTX75J6WZBNC7W4MRMGNYZ/events.json","paper":"https://pith.science/paper/LS6BLTX7"},"agent_actions":{"view_html":"https://pith.science/pith/LS6BLTX75J6WZBNC7W4MRMGNYZ","download_json":"https://pith.science/pith/LS6BLTX75J6WZBNC7W4MRMGNYZ.json","view_paper":"https://pith.science/paper/LS6BLTX7","resolve_alias":"https://pith.science/api/pith-number/resolve?arxiv=2209.14981&json=true","fetch_graph":"https://pith.science/api/pith-number/LS6BLTX75J6WZBNC7W4MRMGNYZ/graph.json","fetch_events":"https://pith.science/api/pith-number/LS6BLTX75J6WZBNC7W4MRMGNYZ/events.json","actions":{"anchor_timestamp":"https://pith.science/pith/LS6BLTX75J6WZBNC7W4MRMGNYZ/action/timestamp_anchor","attest_storage":"https://pith.science/pith/LS6BLTX75J6WZBNC7W4MRMGNYZ/action/storage_attestation","attest_author":"https://pith.science/pith/LS6BLTX75J6WZBNC7W4MRMGNYZ/action/author_attestation","sign_citation":"https://pith.science/pith/LS6BLTX75J6WZBNC7W4MRMGNYZ/action/citation_signature","submit_replication":"https://pith.science/pith/LS6BLTX75J6WZBNC7W4MRMGNYZ/action/replication_record"}},"created_at":"2026-07-05T05:04:04.075922+00:00","updated_at":"2026-07-05T05:04:04.075922+00:00"}