NanoKnow_Benchmark
收藏资源简介:
NanoKnow Benchmark Qrels 是一个预构建的相关性判断数据集,用于将 SQuAD 和 Natural Questions (NQ) 划分为支持和不支持的分割,划分依据是答案是否出现在 nanochat 预训练语料库中。该数据集是 NanoKnow 项目的一部分,旨在测试模型的参数化知识和外部知识依赖能力。数据集包含 SQuAD 和 NQ-Open 的问题,其中 SQuAD 有 10,570 个问题(72% 支持,28% 不支持),NQ-Open 有 3,610 个问题(66% 支持,34% 不支持)。支持的问题包含答案在预训练语料库中的文档标识符和答案偏移量,不支持的问题仅包含问题标识符、问题和答案。数据集文件格式为文本文件,支持的问题包含 qid、question、answer、doc_id 和 answer_offset 字段,不支持的问题包含 qid、question 和 answer 字段。数据集生成采用了 BM25 检索、答案字符串匹配和 LLM 验证的三阶段流程。
NanoKnow Benchmark Qrels is a pre-built relevance judgment dataset that partitions SQuAD and Natural Questions (NQ) into supported and unsupported subsets based on whether the answer appears in the nanochat pre-training corpus. As a component of the NanoKnow project, this dataset aims to assess models' parametric knowledge and external knowledge dependency abilities. The dataset includes questions from both SQuAD and NQ-Open: SQuAD contains 10,570 questions, 72% of which are supported and 28% unsupported; NQ-Open has 3,610 questions, with 66% supported and 34% unsupported. Supported questions come with the document identifier and answer offset of the answer in the pre-training corpus, while unsupported questions only contain the qid, question, and answer fields. The dataset is stored in plain text files: supported entries feature the fields qid, question, answer, doc_id, and answer_offset, whereas unsupported entries include qid, question, and answer. The dataset is generated via a three-stage pipeline encompassing BM25 retrieval, answer string matching, and LLM validation.



