Arena-Write is a writing task dataset for evaluating ultra-long text generation models, released in 2025 by the Singapore University of Technology and Design in collaboration with the Knowledge Engineering Lab of Tsinghua University. The related research papers are as follows: LongWriter-Zero: Mastering Ultra-Long Text Generation via Reinforcement LearningThe aim is to systematically evaluate the comprehensive capabilities of large language models in generating long-form content and complex writing tasks under conditions that closely resemble real-world usage scenarios.

This dataset contains 100 user writing tasks, each consisting of a real-world writing prompt and labeled with the corresponding writing scenario type. The tasks cover various text formats, including social media posts, articles, and reports, and exhibit significant differences in output length, ranging from short text tasks of a few hundred words to long text tasks requiring the generation of over 2,000 words. In addition to the writing prompts, the dataset also provides the generation results of several mainstream baseline models on the same task, supporting comparative evaluation of different model outputs.