Multimodal LLMs for Visualization Reconstruction and Understanding
收藏资源简介:
本数据集由南洋理工大学VIDA实验室创建,旨在帮助机器更好地理解和重建可视化信息。数据集包括2,999个可视化图像,涵盖条形图、折线图和面积图,每个图像都附有相应的编码信息和SimVec表示。SimVec是一种简化的向量格式,能够高效且一致地表示可视化内容。数据集还支持中间推理过程,有助于模型理解可视化图像中的数据。此外,数据集还包含手绘和纸张纹理的旧式可视化,以增强模型在实际场景中的应用能力。
This dataset was developed by the VIDA Lab at Nanyang Technological University (NTU) to enable machines to better understand and reconstruct visual information. It comprises 2,999 visual charts, including bar charts, line charts, and area charts. Each chart is accompanied by corresponding encoding information and SimVec representations. SimVec is a streamlined vector format that can efficiently and consistently depict visual content. The dataset also supports intermediate reasoning processes, which helps models understand the data contained in the visual charts. Furthermore, the dataset includes vintage visualizations with hand-drawn styles and paper textures to enhance the model's applicability in real-world scenarios.
数据集概述:VisNLModel
数据集用途
- 用于训练MiniCPM-V和Qwen-VL两种视觉语言模型
- 支持两种任务类型:
- DataFeat Tasks:包含CoT(Chain-of-Thought)中间推理过程
- SimVec Tasks:图表编码描述任务
数据格式
MiniCPM-V训练数据格式
json { "id": "唯一标识符", "image": "图像路径", "conversations": [ { "role": "user", "content": "<image> 问题文本" }, { "role": "assistant", "content": "回答文本" } ] }
Qwen-VL训练数据格式
json { "id": "唯一标识符", "image": "图像路径", "conversations": [ { "from": "user", "value": "Picture 1: <img>图像路径</img> 问题文本" }, { "from": "assistant", "value": "回答文本" } ] }
数据示例
DataFeat Tasks示例
- 任务:从图表中提取具体数值
- 特点:包含像素到数值的映射计算过程
- 示例问题:"What is the max value of CO2 emissions for Waste Management?"
- 示例回答:包含像素位置和数值映射关系的详细计算过程
SimVec Tasks示例
- 任务:图表编码描述
- 特点:提供图表的向量化表示
- 示例问题:"Describe the encoding of the chart."
- 示例回答:包含坐标轴信息和图表元素的详细向量描述
训练与推理
- 训练脚本:
- MiniCPM-V:
MiniCPM-V/finetune/finetune_lora.sh
- MiniCPM-V:
- 推理脚本:
- MiniCPM-V:
inference_with_MiniCPM.py - Qwen-VL:
inference_with_Qwen.py - GPT-4对比:
inference_with_GPT4o.py
- MiniCPM-V:
评估与结果
- 结果文件:
results.csv - 评估过程可能需要后处理和人工检查
数据获取
- 下载地址:https://huggingface.co/datasets/clda/VisNLModel

- 1Multimodal LLMs for Visualization Reconstruction and Understanding南洋理工大学 · 2025年



