Spark-234K
收藏资源简介:
Spark-234K是一个基于大规模科学文献的骨架引导科学推理数据集,包含约234,000个问答对。该数据集由机器从前沿科学文献中合成生成,采用了一种新颖的骨架引导方法:首先将每篇文献提炼为一个紧凑的推理骨架,保留其核心主张、支持证据、定量关系、假设和边界条件,然后从多个互补推理视角构建具有挑战性且自包含的问题。构建过程基于SCI-BASE语料库(约336万篇论文,涵盖10个学科),经过过滤和学科平衡采样后选取约37万篇前沿论文,再依次通过推理骨架提取、多视角问题生成、答案生成与质量检查、去重与去污染四个阶段。数据集在10个科学学科上平衡分布,其中物理、数学与计算机科学、化学占比最大。质量方面,自包含率达到99.73%,超过93%的问题被归类为L4多步推理或L5研究级推理,人工审核(300样本,跨10学科)显示与专家答案的一致率达90.3%。在Qwen系列模型上进行监督微调时,Spark-234K持续优于规模更大的其他科学推理数据集。该数据集适用于文本生成任务,尤其是科学推理、思维链训练等场景。
Spark-234K is a skeleton-guided scientific reasoning dataset based on large-scale scientific literature, containing approximately 234,000 question-answer pairs. The dataset is synthetically generated by machines from cutting-edge scientific literature using a novel skeleton-guided approach: first, each paper is distilled into a compact reasoning skeleton, retaining its core claims, supporting evidence, quantitative relationships, assumptions, and boundary conditions; then, challenging and self-contained questions are constructed from multiple complementary reasoning perspectives. The construction process is based on the SCI-BASE corpus (approximately 3.36 million papers covering 10 disciplines). After filtering and discipline-balanced sampling, about 370,000 frontier papers are selected, and then processed through four stages: reasoning skeleton extraction, multi-perspective question generation, answer generation and quality check, and deduplication and decontamination. The dataset is balanced across 10 scientific disciplines, with physics, mathematics and computer science, and chemistry accounting for the largest proportions. In terms of quality, the self-containment rate reaches 99.73%, and over 93% of questions are classified as L4 multi-step reasoning or L5 research-level reasoning. Human evaluation (300 samples across 10 disciplines) shows a 90.3% agreement rate with expert answers. When used for supervised fine-tuning on the Qwen series models, Spark-234K consistently outperforms other larger scientific reasoning datasets. This dataset is suitable for text generation tasks, especially for scientific reasoning and chain-of-thought training.
Spark-234K 数据集概述
基本信息
- 数据集名称:Spark-234K
- 语言:英文(单语)
- 许可证:Apache-2.0
- 任务类型:文本生成
- 规模:100K < n < 1M(包含234K条问答对)
- 标注创建者:机器生成
- 数据格式:Parquet文件(Spark-234K-data.parquet,含train split)
数据集简介
Spark-234K(SPARK: Skeleton-Guided Scientific Reasoning)是一个从大规模前沿科学文献中综合生成的高质量科学推理数据集。其核心创新在于不直接从完整论文生成问答对,而是先将每篇论文提炼为紧凑的推理骨架——保留核心主张、支持性证据、定量关系、假设与边界条件——再基于互补的推理视角构建具有挑战性且自包含的问题。该数据集已被EMNLP 2026 Findings接收。
数据构建流程
数据集基于SCI-BASE(涵盖10个学科、约336万篇论文),经筛选和学科平衡采样后选取约37万篇前沿论文,通过以下四个阶段构建:
- 推理骨架提取:将每篇论文提炼为以主张为中心的骨架,捕捉其核心结论、论证步骤、可观测变量和假设。
- 多视角问题生成:从多个互补的推理视角生成问题,以最大化难度和多样性。
- 答案生成与质量检查:在不接触源论文的情况下生成答案,并对每条问答对进行自包含性、来源一致性、答案连贯性和推理基础验证。
- 去重与去污染:按学科进行语义去重,并针对下游基准进行13-gram精确匹配去污染。
学科分布
数据集覆盖10个科学学科,采用学科平衡采样,其中物理学、数学与计算机科学、化学占比最大。
数据质量
- 自包含率:99.73%
- 难度分布:超过93%的问题被归类为L4多步推理或L5研究级推理
- 人工审计:对10个学科300个样本进行分层人工审计,与专家答案的一致率达90.3%
实验效果
在Qwen系列模型上进行SFT训练时,Spark-234K一致优于规模更大的其他科学推理数据集:
| 训练数据 | Qwen3-8B | Qwen3-14B |
|---|---|---|
| 最佳竞品数据集 | 56.36 | 60.45 |
| Spark-234K | 60.83 | 63.88 |
引用信息
如需引用该数据集,可参考论文《SPARK: Skeleton-Guided Reasoning Synthesis from Large-Scale Scientific Literature》,作者包括Yu Li、Wei Li、Xin Gao、Mengyuan Sun、Xiaoyang Wang、Qizhi Pei和Lijun Wu,论文预印本地址为https://arxiv.org/abs/2608.30214。




