{"paper":{"title":"HellaSwag: Can a Machine Really Finish Your Sentence?","license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","headline":"HellaSwag shows state-of-the-art models still fail at commonsense sentence completion that humans solve easily.","cross_cats":[],"primary_cat":"cs.CL","authors_text":"Ali Farhadi, Ari Holtzman, Rowan Zellers, Yejin Choi, Yonatan Bisk","submitted_at":"2019-05-19T23:57:23Z","abstract_excerpt":"Recent work by Zellers et al. (2018) introduced a new task of commonsense natural language inference: given an event description such as \"A woman sits at a piano,\" a machine must select the most likely followup: \"She sets her fingers on the keys.\" With the introduction of BERT, near human-level performance was reached. Does this mean that machines can perform human level commonsense inference?\n  In this paper, we show that commonsense inference still proves difficult for even state-of-the-art models, by presenting HellaSwag, a new challenge dataset. Though its questions are trivial for humans "},"claims":{"count":4,"items":[{"kind":"strongest_claim","text":"Though its questions are trivial for humans (>95% accuracy), state-of-the-art models struggle (<48%).","source":"verdict.strongest_claim","status":"machine_extracted","claim_id":"C1","attestation":"unclaimed"},{"kind":"weakest_assumption","text":"That adversarial filtering produces examples requiring genuine commonsense reasoning rather than merely exploiting specific weaknesses or distributional artifacts in the models used during filtering.","source":"verdict.weakest_assumption","status":"machine_extracted","claim_id":"C2","attestation":"unclaimed"},{"kind":"one_line_summary","text":"HellaSwag dataset shows state-of-the-art models fail commonsense inference tasks that humans solve easily, built via adversarial filtering of distractors.","source":"verdict.one_line_summary","status":"machine_extracted","claim_id":"C3","attestation":"unclaimed"},{"kind":"headline","text":"HellaSwag shows state-of-the-art models still fail at commonsense sentence completion that humans solve easily.","source":"verdict.pith_extraction.headline","status":"machine_extracted","claim_id":"C4","attestation":"unclaimed"}],"snapshot_sha256":"0aba65de350fc7c765ee8f342d88f717b5e1797142ba6879a16035483c3f1f3c"},"source":{"id":"1905.07830","kind":"arxiv","version":1},"verdict":{"id":"4928dd65-3589-403a-bee7-d0205863269c","model_set":{"reader":"grok-4.3"},"created_at":"2026-05-11T02:51:33.151353Z","strongest_claim":"Though its questions are trivial for humans (>95% accuracy), state-of-the-art models struggle (<48%).","one_line_summary":"HellaSwag dataset shows state-of-the-art models fail commonsense inference tasks that humans solve easily, built via adversarial filtering of distractors.","pipeline_version":"pith-pipeline@v0.9.0","weakest_assumption":"That adversarial filtering produces examples requiring genuine commonsense reasoning rather than merely exploiting specific weaknesses or distributional artifacts in the models used during filtering.","pith_extraction_headline":"HellaSwag shows state-of-the-art models still fail at commonsense sentence completion that humans solve easily."},"integrity":{"clean":true,"summary":{"advisory":0,"critical":0,"by_detector":{},"informational":0},"endpoint":"/pith/1905.07830/integrity.json","findings":[],"available":true,"detectors_run":[],"snapshot_sha256":"c28c3603d3b5d939e8dc4c7e95fa8dfce3d595e45f758748cecf8e644a296938"},"references":{"count":19,"sample":[{"doi":"","year":2018,"title":"Yonatan Belinkov and Yonatan Bisk. 2018. Synthetic and natural noise both break neural machine translation. In ICLR. ICLR","work_id":"2f87ef9e-973a-4f33-93e7-a79e927ca32e","ref_index":1,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":2017,"title":"Qian Chen, Xiaodan Zhu, Zhen-Hua Ling, Si Wei, Hui Jiang, and Diana Inkpen. 2017. Enhanced lstm for natural language inference. In Proceedings of the 55th Annual Meeting of the Association for Computa","work_id":"8bd6683b-0e73-4ec0-bb86-391a37e265de","ref_index":2,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":2018,"title":"BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding","work_id":"ed240a10-5b19-406c-baa5-30803f465785","ref_index":3,"cited_arxiv_id":"1810.04805","is_internal_anchor":true},{"doi":"","year":2018,"title":"Max Glockner, Vered Shwartz, and Yoav Goldberg. 2018. Breaking nli systems with sentences that require simple lexical inferences. In Proceedings of the 56th Annual Meeting of the Association for Compu","work_id":"277326a1-be96-4bb7-aab1-64021dbd9013","ref_index":4,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":2013,"title":"Jonathan Gordon and Benjamin Van Durme. 2013. Reporting bias and knowledge acquisition. In Proceedings of the 2013 workshop on Automated knowledge base construction, pages 25--30. ACM","work_id":"1002e502-b170-4aca-bbbd-64a081e2f89b","ref_index":5,"cited_arxiv_id":"","is_internal_anchor":false}],"resolved_work":19,"snapshot_sha256":"fc7576631fb4f5e083bb1553be3225ef4ffe26fe3c0026780127e85c0555c0e6","internal_anchors":2},"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"}