$\"image.png\"$

\n","updatedAt":"2025-02-18T06:45:36.378Z","author":{"_id":"64587be872b60ae7a3817858","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/64587be872b60ae7a3817858/BbdOOxOCEzWTvEpkWp8MM.png","fullname":"Minbyul Jeong","name":"Minbyul","type":"user","isPro":false,"isHf":false,"isHfAdmin":false,"isMod":false,"followerCount":4,"isUserFollowing":false}},"numEdits":0,"identifiedLanguage":{"language":"en","probability":0.8730562925338745},"editors":["Minbyul"],"editorAvatarUrls":["https://cdn-avatars.huggingface.co/v1/production/uploads/64587be872b60ae7a3817858/BbdOOxOCEzWTvEpkWp8MM.png"],"reactions":[],"isReport":false}},{"id":"67b5359115ec269198a734bc","author":{"_id":"63d3e0e8ff1384ce6c5dd17d","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/1674830754237-63d3e0e8ff1384ce6c5dd17d.jpeg","fullname":"Librarian Bot (Bot)","name":"librarian-bot","type":"user","isPro":false,"isHf":false,"isHfAdmin":false,"isMod":false,"followerCount":318,"isUserFollowing":false},"createdAt":"2025-02-19T01:36:17.000Z","type":"comment","data":{"edited":false,"hidden":false,"latest":{"raw":"This is an automated message from the [Librarian Bot](https://huggingface.co/librarian-bots). I found the following papers similar to this paper. \n\nThe following papers were recommended by the Semantic Scholar API \n\n* [IHEval: Evaluating Language Models on Following the Instruction Hierarchy](https://huggingface.co/papers/2502.08745) (2025)\n* [Do LLMs Recognize Your Preferences? Evaluating Personalized Preference Following in LLMs](https://huggingface.co/papers/2502.09597) (2025)\n* [UltraIF: Advancing Instruction Following from the Wild](https://huggingface.co/papers/2502.04153) (2025)\n* [DeepThink: Aligning Language Models with Domain-Specific User Intents](https://huggingface.co/papers/2502.05497) (2025)\n* [Repository-level Code Search with Neural Retrieval Methods](https://huggingface.co/papers/2502.07067) (2025)\n* [Optimization is Better than Generation: Optimizing Commit Message Leveraging Human-written Commit Message](https://huggingface.co/papers/2501.09861) (2025)\n* [A Survey of Personalized Large Language Models: Progress and Future Directions](https://huggingface.co/papers/2502.11528) (2025)\n\n\n Please give a thumbs up to this comment if you found it helpful!\n\n If you want recommendations for any Paper on Hugging Face checkout [this](https://huggingface.co/spaces/librarian-bots/recommend_similar_papers) Space\n\n You can directly ask Librarian Bot for paper recommendations by tagging it in a comment: `@librarian-bot recommend`","html":"

This is an automated message from the Librarian Bot. I found the following papers similar to this paper.

The following papers were recommended by the Semantic Scholar API

Please give a thumbs up to this comment if you found it helpful!

If you want recommendations for any Paper on Hugging Face checkout this Space

You can directly ask Librarian Bot for paper recommendations by tagging it in a comment: \n\n@librarian-bot\n\t recommend

\n","updatedAt":"2025-02-19T01:36:17.114Z","author":{"_id":"63d3e0e8ff1384ce6c5dd17d","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/1674830754237-63d3e0e8ff1384ce6c5dd17d.jpeg","fullname":"Librarian Bot (Bot)","name":"librarian-bot","type":"user","isPro":false,"isHf":false,"isHfAdmin":false,"isMod":false,"followerCount":318,"isUserFollowing":false}},"numEdits":0,"identifiedLanguage":{"language":"en","probability":0.7512053847312927},"editors":["librarian-bot"],"editorAvatarUrls":["https://cdn-avatars.huggingface.co/v1/production/uploads/1674830754237-63d3e0e8ff1384ce6c5dd17d.jpeg"],"reactions":[],"isReport":false}}],"primaryEmailConfirmed":false,"paper":{"id":"2502.11330","authors":[{"_id":"67b42c5632929e97a92dee90","user":{"_id":"64587be872b60ae7a3817858","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/64587be872b60ae7a3817858/BbdOOxOCEzWTvEpkWp8MM.png","isPro":false,"fullname":"Minbyul Jeong","user":"Minbyul","type":"user"},"name":"Minbyul Jeong","status":"claimed_verified","statusLastChangedAt":"2025-02-19T09:04:45.723Z","hidden":false},{"_id":"67b42c5632929e97a92dee91","user":{"_id":"6596a87480a4560b8f9b9532","avatarUrl":"/avatars/33f39ee01c1648f1daca41a40a4964fb.svg","isPro":false,"fullname":"Christopher Cho","user":"ChristopherCho","type":"user"},"name":"Jungho Cho","status":"claimed_verified","statusLastChangedAt":"2025-02-21T09:59:57.458Z","hidden":false},{"_id":"67b42c5632929e97a92dee92","name":"Minsoo Khang","hidden":false},{"_id":"67b42c5632929e97a92dee93","user":{"_id":"6268ec3bd4aa92e53a66d028","avatarUrl":"/avatars/af7fd3b4a49aca8b71617f3f17673227.svg","isPro":false,"fullname":"dawoon jung","user":"noowad","type":"user"},"name":"Dawoon Jung","status":"admin_assigned","statusLastChangedAt":"2025-02-19T15:48:50.413Z","hidden":false},{"_id":"67b42c5632929e97a92dee94","user":{"_id":"6551e62ca12178dd75f4592f","avatarUrl":"/avatars/f4213f9cb95847182bcca3281b1a042c.svg","isPro":false,"fullname":"Teakgyu Hong","user":"tghong","type":"user"},"name":"Teakgyu Hong","status":"admin_assigned","statusLastChangedAt":"2025-02-19T15:48:42.592Z","hidden":false}],"publishedAt":"2025-02-17T01:05:31.000Z","submittedOnDailyAt":"2025-02-18T04:15:36.359Z","title":"System Message Generation for User Preferences using Open-Source Models","submittedOnDailyBy":{"_id":"64587be872b60ae7a3817858","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/64587be872b60ae7a3817858/BbdOOxOCEzWTvEpkWp8MM.png","isPro":false,"fullname":"Minbyul Jeong","user":"Minbyul","type":"user"},"summary":"System messages play a crucial role in interactions with large language\nmodels (LLMs), often serving as prompts to initiate conversations. Through\nsystem messages, users can assign specific roles, perform intended tasks,\nincorporate background information, specify various output formats and\ncommunication styles. Despite such versatility, publicly available data are\noften lack system messages and subject to strict license constraints in the\nindustry field. Manual labeling of publicly available data with system messages\nthat align with user instructions demands significant resources. In view of\nsuch challenges, our work introduces SysGen, a pipeline for generating system\nmessages with better aligned assistant responses from the supervised\nfine-tuning dataset without system messages. Training on SysGen data has\ndemonstrated substantial improvements in the alignment of model responses with\nsystem messages and user instructions, as demonstrated across various\nopen-source models on the Multifacet benchmark, while maintaining minimal\nimpact on other unseen benchmarks such as Open LLM Leaderboard 2. Our\nqualitative analysis highlights the importance of diverse system messages to\nensure better adaptability across different contexts.","upvotes":15,"discussionId":"67b42c5732929e97a92deed7","ai_summary":"SysGen generates system messages from datasets without them, improving alignment of assistant responses with user instructions in large language models.","ai_keywords":[""]},"canReadDatabase":false,"canManagePapers":false,"canSubmit":false,"hasHfLevelAccess":false,"upvoted":false,"upvoters":[{"_id":"644e1b1d9b4e87c31bab0a14","avatarUrl":"/avatars/88bb4c4a67dc8958069e9014f5e73a0b.svg","isPro":false,"fullname":"Michael Barry","user":"MichaelBarryUK","type":"user"},{"_id":"64587be872b60ae7a3817858","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/64587be872b60ae7a3817858/BbdOOxOCEzWTvEpkWp8MM.png","isPro":false,"fullname":"Minbyul Jeong","user":"Minbyul","type":"user"},{"_id":"648eb1eb59c4e5c87dc116e0","avatarUrl":"/avatars/c636cea39c2c0937f01398c94ead5dad.svg","isPro":false,"fullname":"fdsqefsgergd","user":"T-representer","type":"user"},{"_id":"5efbdc4ac3896117eab961a9","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/1602668910270-5efbdc4ac3896117eab961a9.png","isPro":false,"fullname":"Data Mining and Information Systems Lab","user":"dmis-lab","type":"user"},{"_id":"64c75f7f5671d42e0afa86f0","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/64c75f7f5671d42e0afa86f0/3-DSbC90ehSCnzBcRYxPk.jpeg","isPro":false,"fullname":"HyeonwooKim","user":"choco9966","type":"user"},{"_id":"64e5c8e594aa0690321f6b29","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/no-auth/IW5LRzmPcAM-dri8taMN7.png","isPro":false,"fullname":"Yein Park","user":"P-YI","type":"user"},{"_id":"63a4a19e769ff94bc94a4803","avatarUrl":"/avatars/47ca8a30c1d2f23f49001ba09f41f6f0.svg","isPro":false,"fullname":"Face","user":"HugThat","type":"user"},{"_id":"634dffc49b777beec3bc6448","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/1670144568552-634dffc49b777beec3bc6448.jpeg","isPro":false,"fullname":"Zhipeng Yang","user":"svjack","type":"user"},{"_id":"64f4c8739ee58d48e8507e0e","avatarUrl":"/avatars/4be540dfb4a949f37cba2d3c3729fbde.svg","isPro":false,"fullname":"Dmitrii Stoianov","user":"heylimon","type":"user"},{"_id":"6596a87480a4560b8f9b9532","avatarUrl":"/avatars/33f39ee01c1648f1daca41a40a4964fb.svg","isPro":false,"fullname":"Christopher Cho","user":"ChristopherCho","type":"user"},{"_id":"64be260dcf4f379eebb6c4ba","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/64be260dcf4f379eebb6c4ba/HRiQISkB8Ncde4z3H5qyr.jpeg","isPro":false,"fullname":"Jerry","user":"jihoo-kim","type":"user"},{"_id":"66d14ca85b16d47da4af8679","avatarUrl":"/avatars/27871c1374454b0968ac362f12b7ff10.svg","isPro":false,"fullname":"Khang Minsoo","user":"minsooKhang","type":"user"}],"acceptLanguages":["*"],"dailyPaperRank":0}">

Papers

arxiv:2502.11330

System Message Generation for User Preferences using Open-Source Models

Published on Feb 17, 2025

· Submitted by

Minbyul Jeong on Feb 18, 2025

Upvote

Authors:

Minbyul Jeong ,

Jungho Cho ,

Dawoon Jung ,

Teakgyu Hong

Abstract

SysGen generates system messages from datasets without them, improving alignment of assistant responses with user instructions in large language models.

AI-generated summary

System messages play a crucial role in interactions with large language models (LLMs), often serving as prompts to initiate conversations. Through system messages, users can assign specific roles, perform intended tasks, incorporate background information, specify various output formats and communication styles. Despite such versatility, publicly available data are often lack system messages and subject to strict license constraints in the industry field. Manual labeling of publicly available data with system messages that align with user instructions demands significant resources. In view of such challenges, our work introduces SysGen, a pipeline for generating system messages with better aligned assistant responses from the supervised fine-tuning dataset without system messages. Training on SysGen data has demonstrated substantial improvements in the alignment of model responses with system messages and user instructions, as demonstrated across various open-source models on the Multifacet benchmark, while maintaining minimal impact on other unseen benchmarks such as Open LLM Leaderboard 2. Our qualitative analysis highlights the importance of diverse system messages to ensure better adaptability across different contexts.

View arXiv page View PDF Add to collection

Community

Minbyul

Paper author Paper submitter Feb 18, 2025

System messages, also known as initial prompt, play a crucial role in steering LLM behaviors during conversations, shaping their behavior by providing roles, background information, task instructions, and communication styles.

※ Despite their versatility, publicly available datasets rarely include system messages, and those that do often contain only generic ones (e.g., You are a helpful AI assistant). Moreover, in the industry, licensing constraints limit the use of existing datasets and models, making it challenging to develop well-aligned AI assistants.

» To address these challenges, we introduce SYSGEN—a novel pipeline for generating system messages from supervised fine-tuning (SFT) datasets without system messages.
Open-source models fine-tuned with SYSGEN data produce well-aligned assistant responses with both system messages and user instructions, while minimizing performance degradation on the unseen benchmark.

If you’re already leveraging the SFT datasets, we highly recommend applying SYSGEN on top of it for enhanced alignment and usability.