COFOS_data
收藏资源简介:
COFOS_data是用于训练和运行COFOS(一种COF/ROS化学问答模型)的数据集集合。该数据集包含监督微调训练文件以及用于检索增强问答的运行时知识图谱/BM25工件。数据集通过以下方式构建:1) 教师模型蒸馏:使用35B Qwen教师模型重写COF/ROS提示和草稿答案,生成更清晰的学生训练答案;2) RAG风格样本:将问题与知识图谱事实和检索到的证据配对,使学生模型学习基于证据的回答;3) 化学/氧化还原问答:添加辅助的英语化学样本,以加强氧化、还原、电化学和ROS推理能力;4) 风格校正:添加示例使答案更自然,并保留领域区分(如将H2O2视为氧衍生产物而非自由基ROS)。数据集内容分为两部分:训练部分包含四个JSONL格式的监督微调文件(cofos_teacher_distill.jsonl、cofos_rag_sft.jsonl、chem_redox_sft.jsonl、style_correction_sft.jsonl),运行时部分包含知识图谱节点/边/三元组和BM25文档/元数据/索引。该数据集适用于:COF/ROS问答模型的监督微调、训练模型区分主要ROS/产物、次要物种和中间体、训练RAG感知的回答格式、改进围绕ROS和氧衍生产物的一般化学/氧化还原推理、以及使用本地KG/BM25检索数据运行COFOS。需要注意的是,数据集样本包含教师模型重写,应视为监督微调数据而非原始文献证据;专注于COF/ROS和氧化还原问答,并非广泛的化学语料库;可能包含从源记录、检索或教师生成中继承的不完整或噪声标签;不应作为高风险化学安全、医疗、法律或监管决策的唯一依据。
COFOS_data is a dataset collection for training and running COFOS, a COF/ROS chemistry question-answering model. The dataset contains supervised fine-tuning training files as well as runtime knowledge graph/BM25 artifacts for retrieval-augmented question answering. The dataset is constructed through: 1) Teacher model distillation: using a 35B Qwen teacher model to rewrite COF/ROS prompts and draft answers, generating clearer student training answers; 2) RAG-style samples: pairing questions with knowledge graph facts and retrieved evidence to enable the student model to learn evidence-based responses; 3) Chemistry/redox question answering: adding auxiliary English chemistry samples to enhance oxidation, reduction, electrochemistry, and ROS reasoning capabilities; 4) Style correction: adding examples to make answers more natural and preserve domain distinctions (e.g., treating H2O2 as an oxygen-derived product rather than a radical ROS). The dataset content is divided into two parts: the training part includes four JSONL-format supervised fine-tuning files (cofos_teacher_distill.jsonl, cofos_rag_sft.jsonl, chem_redox_sft.jsonl, style_correction_sft.jsonl), and the runtime part includes knowledge graph nodes/edges/triples and BM25 documents/metadata/indexes. The dataset is suitable for: supervised fine-tuning of COF/ROS question-answering models, training models to distinguish major ROS/products, minor species and intermediates, training RAG-aware response formats, improving general chemistry/redox reasoning around ROS and oxygen-derived products, and running COFOS using local KG/BM25 retrieval data. It should be noted that the dataset samples include teacher model rewrites and should be considered as supervised fine-tuning data rather than original literature evidence; it focuses on COF/ROS and redox question answering, not a broad chemistry corpus; it may contain incomplete or noisy labels inherited from source records, retrieval, or teacher generation; it should not be used as the sole basis for high-risk chemical safety, medical, legal, or regulatory decisions.
数据集概述:COFOS_data
- 数据集名称:COFOS_data
- 许可协议:MIT
- 语言:英语
- 标签:化学、共价有机框架(COF)、活性氧物种(ROS)、氧化还原、SFT、RAG、科学问答
- 任务类别:文本生成、问答
构建方式
- 教师蒸馏问答:使用35B参数的Qwen教师模型,将COF/ROS相关的提示和草稿答案重写为更清晰的学生训练答案。
- RAG风格样本:将问题与知识图谱事实和检索到的证据配对,使模型能够学习基于证据的问答方式。
- 化学/氧化还原问答:补充了英语化学样本,以增强氧化、还原、电化学和ROS推理能力。
- 风格校正:添加了示例使答案更加自然,并保留领域区分,如将
H2O2视为氧衍生产物而非自由基ROS。 - 运行时检索数据:为COFOS命令行界面和本地RAG工作流程导出了知识图谱三元组和BM25索引。
数据集内容
- 训练数据 (
train/):cofos_teacher_distill.jsonl:教师蒸馏的COF/ROS问答数据cofos_rag_sft.jsonl:带有上下文/证据的RAG格式问答数据chem_redox_sft.jsonl:化学与氧化还原问答数据style_correction_sft.jsonl:答案风格与校正样本数据
- 运行时数据 (
runtime/):kg/:COF/ROS知识图谱的节点、边、三元组bm25/:BM25文档、元数据和索引文件(非训练样本,用于COFOS CLI和本地项目RAG)
主要用途
- 对COF/ROS问答模型进行监督微调
- 训练模型区分主导ROS/产物、次级物种和中间体
- 训练具有RAG意识的答案格式化,能够利用提供的证据
- 提升关于ROS和氧衍生产物的通用化学/氧化还原推理能力
- 运行COFOS并使用本地的知识图谱/BM25检索数据
局限性
- 样本包含教师模型的重写内容,应作为SFT数据而非原始文献证据使用
- 数据集聚焦于COF/ROS和氧化还原问答,并非广泛的化学语料库
- 运行时BM25证据可能包含文献衍生文本,重新分发或下游发表前需检查来源与许可
- 可能存在不完整或带有噪声的标签,源自源记录、检索或教师生成过程
- 不应将此数据集用作高风险化学安全、医疗、法律或监管决策的唯一依据
相关链接
- 模型: Willlzh/COFOS
- 数据集: Willlzh/COFOS_data
- npm CLI: @lzhzzzzwill/cofos





