{"paper":{"title":"Enhancing Chat Language Models by Scaling High-quality Instructional Conversations","license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","headline":"Scaling AI-generated multi-turn conversations to 1.5 million dialogues produces a fine-tuned LLaMA that outperforms Vicuna.","cross_cats":["cs.AI"],"primary_cat":"cs.CL","authors_text":"Bokai Xu, Bowen Zhou, Maosong Sun, Ning Ding, Shengding Hu, Yujia Qin, Yulin Chen, Zhiyuan Liu, Zhi Zheng","submitted_at":"2023-05-23T16:49:14Z","abstract_excerpt":"Fine-tuning on instruction data has been widely validated as an effective practice for implementing chat language models like ChatGPT. Scaling the diversity and quality of such data, although straightforward, stands a great chance of leading to improved performance. This paper aims to improve the upper bound of open-source models further. We first provide a systematically designed, diverse, informative, large-scale dataset of instructional conversations, UltraChat, which does not involve human queries. Our objective is to capture the breadth of interactions that a human might have with an AI a"},"claims":{"count":4,"items":[{"kind":"strongest_claim","text":"Building upon UltraChat, we fine-tune a LLaMA model to create a powerful conversational model, UltraLLaMA. Our evaluations indicate that UltraLLaMA consistently outperforms other open-source models, including Vicuna, the previously recognized state-of-the-art open-source model.","source":"verdict.strongest_claim","status":"machine_extracted","claim_id":"C1","attestation":"unclaimed"},{"kind":"weakest_assumption","text":"That conversations generated entirely by AI without any human queries can still capture the breadth, coherence, and instructional quality needed to produce measurable gains over prior open-source chat models.","source":"verdict.weakest_assumption","status":"machine_extracted","claim_id":"C2","attestation":"unclaimed"},{"kind":"one_line_summary","text":"UltraChat supplies 1.5 million high-quality multi-turn dialogues that, when used to fine-tune LLaMA, produce UltraLLaMA, which outperforms prior open-source chat models including Vicuna.","source":"verdict.one_line_summary","status":"machine_extracted","claim_id":"C3","attestation":"unclaimed"},{"kind":"headline","text":"Scaling AI-generated multi-turn conversations to 1.5 million dialogues produces a fine-tuned LLaMA that outperforms Vicuna.","source":"verdict.pith_extraction.headline","status":"machine_extracted","claim_id":"C4","attestation":"unclaimed"}],"snapshot_sha256":"710d1919331528367595cc8c28e4f88a12b26bdc3eb5f33175453f7168b100c1"},"source":{"id":"2305.14233","kind":"arxiv","version":1},"verdict":{"id":"4be4ab91-40da-4c60-a141-588a1401e04d","model_set":{"reader":"grok-4.3"},"created_at":"2026-05-15T17:19:38.619498Z","strongest_claim":"Building upon UltraChat, we fine-tune a LLaMA model to create a powerful conversational model, UltraLLaMA. Our evaluations indicate that UltraLLaMA consistently outperforms other open-source models, including Vicuna, the previously recognized state-of-the-art open-source model.","one_line_summary":"UltraChat supplies 1.5 million high-quality multi-turn dialogues that, when used to fine-tune LLaMA, produce UltraLLaMA, which outperforms prior open-source chat models including Vicuna.","pipeline_version":"pith-pipeline@v0.9.0","weakest_assumption":"That conversations generated entirely by AI without any human queries can still capture the breadth, coherence, and instructional quality needed to produce measurable gains over prior open-source chat models.","pith_extraction_headline":"Scaling AI-generated multi-turn conversations to 1.5 million dialogues produces a fine-tuned LLaMA that outperforms Vicuna."},"references":{"count":253,"sample":[{"doi":"","year":null,"title":"Chatgpt: Optimizing language models for dialogue , author=. OpenAI , year=","work_id":"ceca96b6-f726-4375-9ecd-a618818dd235","ref_index":1,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":null,"title":"GPT-4 technical report , author=. arXiv , year=","work_id":"ffcdf3f4-14c8-40b3-b4ba-f95863a04266","ref_index":3,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":2023,"title":"Stanford Center for Research on Foundation Models","work_id":"69b724c0-6bc6-406f-a710-5779cc2425cd","ref_index":4,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":null,"title":"Advances in neural information processing systems , volume=","work_id":"12f5a236-ef7a-4d13-b4de-b51465a6f977","ref_index":6,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":null,"title":"Compacter: Efficient low-rank hypercomplex adapter layers , url =","work_id":"1e0b0a38-cc01-4f8a-88af-b43a76d028dd","ref_index":7,"cited_arxiv_id":"","is_internal_anchor":false}],"resolved_work":253,"snapshot_sha256":"b2f9963ccc66510d58c8ba06b2ff4a484ae289dbf3d9077a1eed30fef0926e0b","internal_anchors":15},"formal_canon":{"evidence_count":2,"snapshot_sha256":"216d8e1ee014aa0646de0cd14051631c255f8505af1955361ba88631b13cb39e"},"author_claims":{"count":0,"strong_count":0,"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"builder_version":"pith-number-builder-2026-05-17-v1"}