Research map / Large language models / Language, safety and interpretability
LMSYS-Chat-1M: A Large-Scale Real-World LLM Conversation Dataset
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Introduces LMSYS-Chat-1M, a dataset of one million real-user conversations with 25 LLMs collected from the Vicuna demo and Chatbot Arena, and demonstrates four use cases (moderation, safety benchmark, instruction tuning, challenging benchmark prompts).
Problem
Real-world human-LLM interaction data is scarce because hosting multiple LLMs is costly and commercial vendors rarely release their user logs, limiting research on user behavior, safety, and LLM capabilities.
Approach
A large, publicly released corpus of in-the-wild conversations spanning 25 LLMs (open and proprietary) collected via a free gamified platform, with OpenAI moderation tags, plus demonstrations of how it can be used for moderation, jailbreak safety evaluation, instruction tuning, and benchmark construction.
Key results
- Vicuna-moderator-7B reaches 0.70 one-shot micro-F1 on the moderation benchmark, matching GPT-4 (0.69 one-shot, 0.71 zero-shot) and beating GPT-3.5-Turbo (0.45 zero-shot).
- HighQuality-7B scores 47.7 on MMLU and 6.03 on MT-Bench versus Vicuna-7B-v1.5 at 49.8 and 6.17, while Upvote-7B reaches 45.0 and 5.86.
- On the 50-prompt jailbreak safety benchmark, Llama-2-13B-chat has a 16% success rate and Alpaca-13B 74%, with GPT-4 at 34%.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Fine-tune Vicuna-7B on moderation examples from LMSYS-Chat-1M | Vicuna-7B · 5-category content moderation benchmark, zero-shot | Micro-F1: 0.35 -> 0.65 | Helped |
| Add a one-shot example (ablation) | Vicuna-moderator-7B · 5-category content moderation benchmark | Micro-F1: 0.65 -> 0.70 | Helped |
| Add a one-shot example (ablation) | GPT-4 · 5-category content moderation benchmark | Micro-F1: 0.71 -> 0.69 | Hurt |
| Add a one-shot example (ablation) | GPT-3.5-Turbo · 5-category content moderation benchmark | Micro-F1: 0.45 -> 0.64 | Helped |
| Add a one-shot example (ablation) | Vicuna-7B · 5-category content moderation benchmark | Micro-F1: 0.35 -> 0.50 | Helped |
| Add a one-shot example (ablation) | Claude-2 · 5-category content moderation benchmark | Micro-F1: 0.32 -> 0.30 | Hurt |
| Add a one-shot example (ablation) | Llama-2-7B-chat · 5-category content moderation benchmark | Micro-F1: 0.00 -> 0.01 | Helped |
| Fine-tune Llama2-7B on 45K conversations from OpenAI and Anthropic models | Llama2-7B · MMLU (5-shot) | MMLU: 42.4 -> 47.7 | Helped |
| Fine-tune Llama2-7B on 45K conversations from OpenAI and Anthropic models | Llama2-7B · MT-Bench | MT-Bench score: 3.95 -> 6.03 | Helped |
| Use user-voted conversations from open models instead of the high-quality distilled subset (ablation) | Llama2-7B instruction fine-tuning · MMLU (5-shot) | MMLU: 47.7 -> 45.0 | Hurt |
| Use user-voted conversations from open models instead of the high-quality distilled subset (ablation) | Llama2-7B instruction fine-tuning · MT-Bench | MT-Bench score: 6.03 -> 5.86 | Hurt |
| Select prompts scoring >8 rather than prompts scoring <2 (ablation) | GPT-3.5-Turbo prompt-rating method · Top-50 vs Bottom-50 benchmark prompts | GPT-4 win rate against GPT-3.5-Turbo: 22% -> 52% | Helped |
Surprising findings
- OpenAI moderation API flagged only about 5% of conversations, and its recall is likely low, so unsafe content is probably more prevalent than flagged.
- Llama-2-7B-chat refused nearly all moderation tasks, scoring 0.00 zero-shot micro-F1.
Limitations
- Biased user distribution, mostly LLM hobbyists and researchers rather than general population.
- Contains repeated, low-quality, and template-generated data since no filtering was applied; no human preference annotations included.
- Moderation and topic analyses rely on automated tools (OpenAI moderation API, GPT-4 summaries) with uncertain recall.
Experimental setup
Datasets: LMSYS-Chat-1M ShareGPT MMLU MT-Bench
Benchmarks: LMSYS-Chat-1M moderation eval set (110 toxic messages) Jailbreak safety benchmark (50 conversations) MMLU MT-Bench Arena-Hard-200
Metrics: micro-F1 jailbreak success rate MMLU accuracy MT-Bench score win rate
Baselines: GPT-4 GPT-3.5-Turbo Claude-2 Llama-2-7B-chat Vicuna-7B-v1.5 MT-Bench OpenAI text-moderation-latest
Builds on: Chatbot Arena Vicuna ShareGPT OpenAI Moderation API
Closest papers by meaning
- WildChat-50M: A Deep Dive Into the Role of Synthetic Data in Post-Training (ICML 2025 · similarity 0.91)
- Chatbot Arena: An Open Platform for Evaluating LLMs by Human Preference (ICML 2024 · similarity 0.90)
- MixEval: Deriving Wisdom of the Crowd from LLM Benchmark Mixtures (NeurIPS 2024 · similarity 0.90)
- A Statistical Framework for Ranking LLM-based Chatbots (ICLR 2025 · similarity 0.89)
- WildBench: Benchmarking LLMs with Challenging Tasks from Real Users in the Wild (ICLR 2025 · similarity 0.89)
- LongMemEval: Benchmarking Chat Assistants on Long-Term Interactive Memory (ICLR 2025 · similarity 0.87)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.