{"paper":{"title":"Text and Code Embeddings by Contrastive Pre-Training","license":"http://creativecommons.org/licenses/by/4.0/","headline":"Contrastive pre-training on unsupervised data at scale produces high-quality embeddings for text and code that excel at classification and semantic search.","cross_cats":["cs.LG"],"primary_cat":"cs.CL","authors_text":"Alec Radford, Arvind Neelakantan, Boris Power, Chris Hallacy, David Schnurr, Felipe Petroski Such, Girish Sastry, Gretchen Krueger, Jerry Tworek, Jesse Michael Han, Joanne Jang, Johannes Heidecke, Jong Wook Kim, Kenny Hsu, Lilian Weng, Madeleine Thompson, Nikolas Tezak, Peter Welinder, Pranav Shyam, Qiming Yuan, Raul Puri, Tabarak Khan, Tao Xu, Toki Sherbakov, Tyna Eloundou Nekoul","submitted_at":"2022-01-24T23:36:20Z","abstract_excerpt":"Text embeddings are useful features in many applications such as semantic search and computing text similarity. Previous work typically trains models customized for different use cases, varying in dataset choice, training objective and model architecture. In this work, we show that contrastive pre-training on unsupervised data at scale leads to high quality vector representations of text and code. The same unsupervised text embeddings that achieve new state-of-the-art results in linear-probe classification also display impressive semantic search capabilities and sometimes even perform competit"},"claims":{"count":4,"items":[{"kind":"strongest_claim","text":"contrastive pre-training on unsupervised data at scale leads to high quality vector representations of text and code. The same unsupervised text embeddings that achieve new state-of-the-art results in linear-probe classification also display impressive semantic search capabilities and sometimes even perform competitively with fine-tuned models.","source":"verdict.strongest_claim","status":"machine_extracted","claim_id":"C1","attestation":"unclaimed"},{"kind":"weakest_assumption","text":"That the contrastive objective applied to unsupervised pairs at scale captures semantic similarity in a way that generalizes beyond the specific benchmarks used and is not primarily driven by model scale or data volume alone.","source":"verdict.weakest_assumption","status":"machine_extracted","claim_id":"C2","attestation":"unclaimed"},{"kind":"one_line_summary","text":"Contrastive pre-training on unsupervised data at scale creates text and code embeddings that set new state-of-the-art results on classification and semantic search benchmarks.","source":"verdict.one_line_summary","status":"machine_extracted","claim_id":"C3","attestation":"unclaimed"},{"kind":"headline","text":"Contrastive pre-training on unsupervised data at scale produces high-quality embeddings for text and code that excel at classification and semantic search.","source":"verdict.pith_extraction.headline","status":"machine_extracted","claim_id":"C4","attestation":"unclaimed"}],"snapshot_sha256":"7a45acb93c6806ced9b29b0b9a3c746f6432c69f9b7f1646ace5df0a5b8b306f"},"source":{"id":"2201.10005","kind":"arxiv","version":1},"verdict":{"id":"c4457ccf-07e8-4bfc-8b81-17a68fd2ec52","model_set":{"reader":"grok-4.3"},"created_at":"2026-05-15T19:20:15.627353Z","strongest_claim":"contrastive pre-training on unsupervised data at scale leads to high quality vector representations of text and code. The same unsupervised text embeddings that achieve new state-of-the-art results in linear-probe classification also display impressive semantic search capabilities and sometimes even perform competitively with fine-tuned models.","one_line_summary":"Contrastive pre-training on unsupervised data at scale creates text and code embeddings that set new state-of-the-art results on classification and semantic search benchmarks.","pipeline_version":"pith-pipeline@v0.9.0","weakest_assumption":"That the contrastive objective applied to unsupervised pairs at scale captures semantic similarity in a way that generalizes beyond the specific benchmarks used and is not primarily driven by model scale or data volume alone.","pith_extraction_headline":"Contrastive pre-training on unsupervised data at scale produces high-quality embeddings for text and code that excel at classification and semantic search."},"references":{"count":28,"sample":[{"doi":"","year":null,"title":"Evaluating Large Language Models Trained on Code","work_id":"042493e9-b26f-4b4e-bbde-382072ca9b08","ref_index":1,"cited_arxiv_id":"2107.03374","is_internal_anchor":true},{"doi":"","year":null,"title":"SentEval: An evaluation toolkit for universal sentence representations","work_id":"9ca81cef-98b4-4e74-ae6c-d2f5977db107","ref_index":2,"cited_arxiv_id":"1803.05449","is_internal_anchor":true},{"doi":"","year":2005,"title":"Cert: Contrastive self-supervised learning for language understanding","work_id":"c8bee491-a7cb-4564-bb24-4531a5283b58","ref_index":3,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":null,"title":"doi:10.48550/ARXIV.2109.10086","work_id":"b92b66c0-5a02-4966-91fe-a2935c54d59b","ref_index":4,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":2002,"title":"REALM: Retrieval-Augmented Language Model Pre-Training","work_id":"a397ddf8-b0b7-4e32-9d59-fb6ea67ac287","ref_index":5,"cited_arxiv_id":"2002.08909","is_internal_anchor":true}],"resolved_work":28,"snapshot_sha256":"d03ceb661eecf5b6a6f99f0a9a6f210b2b84550cc3bcbf2496bf55ebdcae9aa6","internal_anchors":15},"formal_canon":{"evidence_count":2,"snapshot_sha256":"c9e6d50bd2ee03b93f027ef512c23efb7197b14cea0bae96c746580f8469dbc9"},"author_claims":{"count":0,"strong_count":0,"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"builder_version":"pith-number-builder-2026-05-17-v1"}