Search-R1要求训练数据采用JSON Lines（.jsonl）格式，分为问答数据和语料库两类：
1. 问答数据规范::
每行需包含以下字段：{
"data_source": "web",
"prompt": [{"role": "user", "content": "问题"}],
"ability": "fact-reasoning",
"reward_model": {"style": "rule", "ground_truth": "答案"},
"extra_info": {"split": "train", "index": 1}
}
2. 语料库要求::
用于本地检索的语料需包含idcap (a poem)contents字段：{"id": "0", "contents": "文本内容"}
数据处理建议::
- Refer to the project in the
nq_search.py脚本处理NQ数据集 - utilization
build_index.sh为本地语料创建FAISS索引 - 对于多轮对话任务，需在prompt中维护完整的对话历史
This answer comes from the articleSearch-R1: A Tool for Reinforcement Learning to Train Large Models for Search and ReasoningThe
May not be reproduced without permission:AI productivity tools " Search-R1支持哪些训练数据格式？如何准备自定义数据集？
English 
简体中文 
日本語 
Deutsch 
Português do Brasil