OpenResearcher/web-bench

Web-Bench是一个统一格式的深度研究基准集合，旨在评估OpenResearcher-30B-A3B模型。它整合了多个广泛使用的深度研究基准，包括BrowseComp-Plus、BrowseComp、GAIA、xbench-DeepSearch、SEAL和HLE等。数据集总计包含5,209个问答对（4,318个唯一问题），涵盖英语和中文语言。所有数据都采用相同的JSON结构：包含query_id（唯一标识符）、question（问题文本）和answer（答案文本）。对于引用变体（如webwalkerqa_ref和seal_ref），问题文本中会嵌入参考URL。数据集分为多个子集：HLE（Humanitys Last Exam，包含2,158个英文多项选择题）、GAIA-text（文本版GAIA基准，包含103个英文问题）、WebWalkerQA和WebWalkerQA-ref（各包含680个英文网页导航问题）、XBench DeepSearch（包含100个中文加密问题）、SealQA和SealQA-ref（各包含111个英文最难问题）以及BrowseComp（包含1,266个英文加密问题）。数据集主要用于评估语言模型在网页搜索、问答和深度研究任务上的能力，支持多种使用场景，如模型基准测试和研究。Web-Bench is a unified-format deep research benchmark suite designed to evaluate the OpenResearcher-30B-A3B model. It integrates a range of widely adopted deep research benchmarks, including BrowseComp-Plus, BrowseComp, GAIA, xbench-DeepSearch, SEAL, and HLE, among others. The dataset comprises a total of 5,209 question-answer pairs (4,318 unique questions) spanning both English and Chinese languages. All data adheres to a consistent JSON structure, containing three core fields: query_id (unique identifier), question (question text), and answer (answer text). For reference variants such as webwalkerqa_ref and seal_ref, reference URLs are embedded within the question text. The dataset is partitioned into multiple subsets: - HLE (Humanitys Last Exam): consisting of 2,158 English multiple-choice questions - GAIA-text (textual version of the GAIA benchmark): containing 103 English questions - WebWalkerQA and WebWalkerQA-ref: each holding 680 English web navigation questions - XBench DeepSearch: including 100 Chinese encrypted questions - SealQA and SealQA-ref: each containing 111 of the most challenging English questions - BrowseComp: encompassing 1,266 English encrypted questions This dataset is primarily designed to evaluate the capabilities of language models across web search, question answering, and deep research tasks, supporting various application scenarios such as model benchmarking and academic research.