ali5341/scifact-chat-format

SciFact Chat-Format数据集是基于`allenai/scifact`原始数据集的聊天格式准备，用于监督微调（SFT）。原始数据集专注于科学声明验证，提供了支持（SUPPORT）、反驳（CONTRADICT）和证据不足（NOT ENOUGH INFO）三种标签。数据集包含了约1.4K个带有证据摘要的声明。数据准备过程中，每个声明-证据集被转换为一行数据，包括用户指令、声明、标题和索引摘要，以及助手的结构化响应（标签、理由句子ID和解释）。数据集格式为OpenAI风格的`messages`格式，适用于指令调优对话。

The SciFact Chat-Format dataset is a chat-format preparation of the original `allenai/scifact` dataset for supervised fine-tuning (SFT). The original dataset focuses on scientific claim verification, providing labels for SUPPORT, CONTRADICT, and NOT ENOUGH INFO. It includes approximately 1.4K claims with evidence-containing abstracts. During preparation, each claim-evidence set is converted into a single row of data, including user instructions, claims, titles, and indexed abstracts, as well as structured assistant responses (label, rationale sentence IDs, and explanation). The dataset is formatted in the OpenAI-style `messages` format, suitable for instruction-tuning conversations.