ali5341/qasper-chat-format
收藏资源简介:
该数据集是QASPER数据集的聊天格式准备版本,用于监督微调(SFT)。数据集包含训练、验证和统计文件,以及一个准备脚本。原始数据集专注于科学NLP论文的问题回答和证据选择,包含5,049个问题和1,585篇论文,具有多种答案类型(自由形式、抽取式、是/否、无法回答)和证据注释。数据集的准备过程包括每个(论文,问题)的最佳注释,答案归一化优先级(自由形式 > 是/否 > 抽取式 > 无法回答),以及混合上下文模式(仅证据或全文)。每个JSONL行包含用户和助手的消息,以及元数据(ID、答案类型、上下文模式、证据计数)。
This dataset is a chat-format preparation of QASPER for supervised fine-tuning (SFT). It includes training, validation, and statistics files, as well as a preparation script. The original dataset focuses on question answering and evidence selection for scientific NLP papers, containing 5,049 questions over 1,585 papers with multiple answer types (free-form, extractive, yes/no, unanswerable) and evidence annotations. The preparation process involves best available annotations per (paper, question), answer normalization priority (free-form > yes/no > extractive > unanswerable), and mixed context mode (evidence-only or full-text). Each JSONL row contains messages from user and assistant, along with metadata (IDs, answer type, context mode, evidence count).




