lamm-mit/graph-reasoning-messages-11K

Graph Reasoning (Messages)是一个用于训练和评估语言模型中图推理/结构化推理行为的聊天消息集合。数据集包含11,024个训练样本，每个样本都是OpenAI风格的聊天对话格式，由一系列{role, content}消息组成。助手回复包含显式中间结构（如<think>, <graph>, <graph_json>等标签），可用于监督图提取、结构化分解、推理跟踪格式化等任务。数据集主要用于监督微调(SFT)、评估图提取保真度和工具实验（如将自然语言转换为图）等用途。数据格式为JSON列表，包含role（user/assistant等）和content（消息文本）字段。

Graph Reasoning (Messages) is a collection of chat messages designed for training and evaluating graph-native/structured reasoning behaviors in language models. The dataset contains 11,024 training examples, each represented as an ordered list of {role, content} messages (OpenAI-style chat format). Assistant responses include explicit intermediate structure (e.g., <think>, <graph>, <graph_json> tags) that can be used to supervise: graph extraction, structured decomposition, reasoning trace formatting, and synthesis from structured artifacts. Primary intended uses include supervised fine-tuning (SFT), evaluation of graph extraction fidelity, and tooling experiments (e.g., converting natural language to graphs). The data format consists of JSON-like message lists with role (user/assistant/etc.) and content (message text) fields.