ChartCoF
收藏资源简介:
ChartCoF数据集是由微软亚洲研究院创建的高质量推理数据集,包含超过19种图表类型。该数据集旨在为细粒度分析和模型微调提供多样化的复杂推理问答对。通过程序化的函数链探索生成多样化的推理路径,并翻译成自然语言的形式,确保了问题的精确性和多样性。数据集的创建过程避免了依赖于大型模型,使得推理数据生成更加高效。ChartCoF数据集的应用领域是图表理解,旨在解决多模态大型语言模型在处理复杂图表推理任务时的性能不足问题。
The ChartCoF dataset is a high-quality reasoning dataset developed by Microsoft Research Asia, encompassing over 19 types of charts. Its core objective is to provide diverse complex reasoning question-answer pairs for fine-grained analysis and model fine-tuning. Diverse reasoning paths are explored and generated through a chain of procedural functions, then translated into natural language, ensuring the accuracy and diversity of the questions. The dataset creation process avoids relying on large-scale models, making the generation of reasoning data more efficient. Focused on the field of chart understanding, the ChartCoF dataset aims to address the performance shortcomings of multimodal large language models when handling complex chart reasoning tasks.

- 1Chain of Functions: A Programmatic Pipeline for Fine-Grained Chart Reasoning Data微软亚洲研究院 · 2025年



