DinoDS/retrieval_grounding

该数据集是一个专注于检索-接地的预览数据集，基于Dino Data的四个能力切片构建：搜索触发检测、接地搜索集成、历史搜索触发和历史搜索集成。目标是训练或检查助手在两个相关问题上的行为：1) 决定何时需要检索或历史查找；2) 生成基于提供证据或先前线程上下文的接地回答。数据集包含80行数据，分为训练集（72行）、验证集（4行）和测试集（4行），语言为英语。每行是一个扁平化的助手训练示例，包含任务和路由元数据。重要列包括sample_id、source_lane、user_message、assistant_response等。数据集可用于检索触发分类实验、接地答案微调、历史感知助手行为研究等。

This dataset is a focused retrieval-grounding preview built from four Dino Data capability slices: search trigger detection, grounded search integration, history search trigger, and history search integration. The goal is to train or inspect assistant behavior around two connected problems: 1) deciding when retrieval or history lookup is needed, and 2) generating answers that stay grounded to supplied evidence or prior thread context. The dataset contains 80 rows, divided into train (72 rows), validation (4 rows), and test (4 rows) sets, with English as the language. Each row is a flattened assistant-training example with task and routing metadata. Important columns include sample_id, source_lane, user_message, assistant_response, etc. The dataset can be used for retrieval-trigger classification experiments, grounded answer fine-tuning, history-aware assistant behavior studies, etc.