Forival/LiveBrowseComp
收藏资源简介:
LiveBrowseComp数据集是一个用于评估搜索代理能力的基准数据集,包含335个复杂推理问题,覆盖历史、地理、影视、体育、政治和学术等多个领域。每个问题以JSONL格式存储,包括问题ID(idx)、问题描述(problem)和参考答案(answer)。问题设计为需要多跳推理和实时搜索才能解答,旨在测试代理在浏览和验证信息时的性能。数据集中的problem和answer字段经过XOR加密处理,以防止数据泄露污染大型语言模型的训练语料库,加密方法基于BrowseComp(OpenAI, 2025)方案,使用特定金丝雀字符串作为密码。解密后,问题示例显示为涉及南美国家电视行业的复杂描述。该数据集由HuiMing Fan等人于2026年发布,引用自相关研究论文。
The LiveBrowseComp dataset is a benchmark for evaluating search agents capabilities, consisting of 335 complex reasoning questions spanning multiple domains such as history, geography, film and television, sports, politics, and academia. Each question is stored in JSONL format, including a question ID (idx), problem description (problem), and reference answer (answer). The questions are designed to require multi-hop reasoning and live search to answer, aiming to test agents performance in browsing and verifying information. The problem and answer fields in the dataset are encrypted using an XOR cipher scheme to prevent data leakage from contaminating large language model training corpora, based on the BrowseComp (OpenAI, 2025) method with a specific canary string as the password. After decryption, an example question involves a complex description related to the television industry in a South American country. This dataset was published by HuiMing Fan et al. in 2026 and cited from relevant research papers.




