Tevatron/browsecomp-plus-md-toc-gpt5.4-nano
收藏资源简介:
该数据集是BrowseComp-Plus 100k语料库的一个官方格式变体,用于RISE Agent实验。它保留了原始语料库的行数(100,195行)、文档ID、URL和列名,但将每个文档的text字段替换为结构化版本,增加了生成的目录(TOC)和章节标题(使用##标记)。结构化文本通过GPT-5.4-nano模型生成章节提议,并经过后处理器验证和实现,确保原始文本不被修改,仅添加导航辅助。数据集旨在用于受控的代理搜索和检索实验,其中语料库与BrowseComp-Plus对齐,但文档提供轻量级导航结构以支持行范围阅读。数据集文件为Parquet格式,包含docid、text和url三列,可用于加载和检索任务。
This dataset is a drop-in, official-format variant of the BrowseComp-Plus 100k corpus used in RISE Agent experiments. It keeps the same row count (100,195 rows), document ids, URLs, and column names as the original BrowseComp-Plus corpus, but replaces each documents text field with a structured version that adds a generated table of contents (TOC) and section headings (marked with ##). The structured text is generated using a GPT-5.4-nano model for section proposals, validated and materialized by a deterministic postprocessor, preserving the original text while adding navigation aids. The dataset is intended for controlled agentic-search and retrieval experiments where the corpus remains aligned with BrowseComp-Plus, but documents expose a lightweight navigation structure for line-range reading. The dataset is in Parquet format with three columns: docid, text, and url, suitable for loading and retrieval tasks.




