DeR2
收藏资源简介:
DeR2是由字节跳动和M-A-P联合开发的科学推理基准测试数据集,旨在解耦检索与推理能力评估。该数据集包含从2023-2025年理论论文中提取的冻结文档库(平均每实例6.5篇),包含专家标注的概念集和验证过的思维链依据。通过四重评估机制(纯指令、概念集、相关文档集、全文档集)实现细粒度错误归因,重点解决前沿科学问题中多步推导、噪声过滤和证据合成等核心难点,为AI模型在科研场景下的证据驱动推理能力提供标准化评估框架。
DeR2 is a scientific reasoning benchmark dataset jointly developed by ByteDance and M-A-P, which aims to decouple the evaluation of retrieval and reasoning capabilities. This dataset contains a frozen document corpus extracted from theoretical papers published between 2023 and 2025, with an average of 6.5 documents per instance, as well as expert-annotated concept sets and verified chain-of-thought evidence. It enables fine-grained error attribution through a four-fold evaluation mechanism: pure instruction, concept set, relevant document set, and full document set, focusing on addressing core challenges in cutting-edge scientific problems such as multi-step deduction, noise filtering, and evidence synthesis, providing a standardized evaluation framework for evidence-driven reasoning capabilities of AI models in scientific research scenarios.




