ORBIT
收藏资源简介:
ORBIT是由滑铁卢大学团队开发的开放网络推理数据集,包含2万条需4-5步推理的复杂查询-答案对,覆盖15个领域(如科技、医学、影视等)。该数据集通过四阶段自动化框架构建:种子创建(基于维基百科分类)、问答生成(DeepSeek-V3.1模型)、自验证和外部验证(Qwen3-4B与GPT-OSS-120B双重审核),最终确保答案可通过网络搜索验证。其创新性在于零预置条件、低成本生成,专为训练小型搜索智能体(如4B参数模型)解决复杂多跳检索问题而设计,在维基百科QA任务中表现优于同类基准9%准确率。
ORBIT is an open web reasoning dataset developed by the team from the University of Waterloo. It contains 20,000 complex query-answer pairs requiring 4 to 5 steps of reasoning, covering 15 domains such as technology, medicine, film and television, and others. This dataset is constructed via a four-stage automated framework: seed creation (based on Wikipedia categories), question-answer generation (using the DeepSeek-V3.1 model), self-verification and external verification (dual audit by Qwen3-4B and GPT-OSS-120B), which ultimately ensures that all answers can be verified through web searches. Its innovative features include zero preconditions and low-cost generation, and it is specifically designed for training small-scale search AI agents (e.g., 4B-parameter models) to solve complex multi-hop retrieval problems. It outperforms comparable benchmarks by 9% in accuracy on the Wikipedia QA task.




