{"paper":{"title":"Magpie: Alignment Data Synthesis from Scratch by Prompting Aligned LLMs with Nothing","license":"http://creativecommons.org/licenses/by/4.0/","headline":"Prompting aligned LLMs like Llama-3-Instruct with only left-side conversation templates produces millions of realistic user queries and responses for alignment training.","cross_cats":["cs.AI"],"primary_cat":"cs.CL","authors_text":"Bill Yuchen Lin, Fengqing Jiang, Luyao Niu, Radha Poovendran, Yejin Choi, Yuntian Deng, Zhangchen Xu","submitted_at":"2024-06-12T17:52:30Z","abstract_excerpt":"High-quality instruction data is critical for aligning large language models (LLMs). Although some models, such as Llama-3-Instruct, have open weights, their alignment data remain private, which hinders the democratization of AI. High human labor costs and a limited, predefined scope for prompting prevent existing open-source data creation methods from scaling effectively, potentially limiting the diversity and quality of public alignment datasets. Is it possible to synthesize high-quality instruction data at scale by extracting it directly from an aligned LLM? We present a self-synthesis meth"},"claims":{"count":4,"items":[{"kind":"strongest_claim","text":"Our results indicate that in some tasks, models fine-tuned with Magpie perform comparably to the official Llama-3-8B-Instruct, despite the latter being enhanced with 10 million data points through supervised fine-tuning (SFT) and subsequent feedback learning.","source":"verdict.strongest_claim","status":"machine_extracted","claim_id":"C1","attestation":"unclaimed"},{"kind":"weakest_assumption","text":"The generated user queries produced by prompting with left-side templates are sufficiently diverse, realistic, and representative of real user needs to support effective alignment after filtering.","source":"verdict.weakest_assumption","status":"machine_extracted","claim_id":"C2","attestation":"unclaimed"},{"kind":"one_line_summary","text":"Magpie synthesizes 300K high-quality alignment instructions from Llama-3-Instruct via auto-regressive prompting on partial templates, enabling fine-tuned models to match official instruct performance on AlpacaEval, ArenaHard, and WildBench.","source":"verdict.one_line_summary","status":"machine_extracted","claim_id":"C3","attestation":"unclaimed"},{"kind":"headline","text":"Prompting aligned LLMs like Llama-3-Instruct with only left-side conversation templates produces millions of realistic user queries and responses for alignment training.","source":"verdict.pith_extraction.headline","status":"machine_extracted","claim_id":"C4","attestation":"unclaimed"}],"snapshot_sha256":"da8a76606000b613e379f3e979ba4b857811040014e2997dc7d20ee7c283bb73"},"source":{"id":"2406.08464","kind":"arxiv","version":2},"verdict":{"id":"4a54f3eb-5e78-4e05-9cc7-81fe03eab948","model_set":{"reader":"grok-4.3"},"created_at":"2026-05-16T06:54:57.162248Z","strongest_claim":"Our results indicate that in some tasks, models fine-tuned with Magpie perform comparably to the official Llama-3-8B-Instruct, despite the latter being enhanced with 10 million data points through supervised fine-tuning (SFT) and subsequent feedback learning.","one_line_summary":"Magpie synthesizes 300K high-quality alignment instructions from Llama-3-Instruct via auto-regressive prompting on partial templates, enabling fine-tuned models to match official instruct performance on AlpacaEval, ArenaHard, and WildBench.","pipeline_version":"pith-pipeline@v0.9.0","weakest_assumption":"The generated user queries produced by prompting with left-side templates are sufficiently diverse, realistic, and representative of real user needs to support effective alignment after filtering.","pith_extraction_headline":"Prompting aligned LLMs like Llama-3-Instruct with only left-side conversation templates produces millions of realistic user queries and responses for alignment training."},"references":{"count":126,"sample":[{"doi":"","year":null,"title":"and Stoica, Ion and Xing, Eric P","work_id":"cb4b41f6-6d60-4db4-a4d1-6c5bb7899473","ref_index":5,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":null,"title":"Interpretable Preferences via Multi-Objective Reward Modeling and Mixture-of-Experts , author=. EMNLP , year=","work_id":"77415ed4-7f1f-44b1-bd55-58146dac1904","ref_index":6,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":2022,"title":"Proceedings of the 2022 ACM Conference on Fairness, Accountability, and Transparency , pages=","work_id":"bd104d56-1e1d-4622-9852-13b7c597c1ae","ref_index":8,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":2018,"title":"Improving language understanding by generative pre-training , author=. 2018 , publisher=","work_id":"3bddebd1-6efa-4e33-b5e0-8e75c670b486","ref_index":9,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":2023,"title":"International Conference on Machine Learning , pages=","work_id":"e6670a94-9685-42c2-84ec-6accc28bc8a4","ref_index":11,"cited_arxiv_id":"","is_internal_anchor":false}],"resolved_work":126,"snapshot_sha256":"f93013ac98f5fddef0513f65b115c13598a8e648edab6b1abca9114f378a3901","internal_anchors":22},"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"}