AIAny
Icon for item

ChartGalaxy

Provides 1.7M+ synthetic and real infographic charts paired with their tabular data for training and evaluating multimodal models on infographic understanding, chart-to-table extraction, chart code generation, and example-based chart synthesis.

Introduction

Most chart datasets focus on plain plots; infographic charts combine visual motifs, icons and rich textual annotations that break assumptions used by many vision–language models. ChartGalaxy addresses this gap by delivering large-scale, design-aware chart data that ties rendered infographic images to the underlying tables and extracted layout templates, so models can learn both visual design cues and the exact data-to-visual mapping.

What Sets It Apart
  • Scale + paired supervision: a multi-million–sample collection of synthetic and real infographic charts where each image is paired with the tabular data that generated it — enabling direct chart↔table supervision for VQA and data extraction tasks, and objective evaluation of code-generation outputs.
  • Design-grounded synthesis: templates, chart types, and style variations are induced from real infographic designs and then used in a human-in-the-loop pipeline to create diverse synthetic charts — so the dataset preserves real-world layout diversity while scaling to millions of examples.
  • Multi-task utility: includes VQA-style QA pairs, layout templates, chart code benchmarks and example-based generation setups, making it useful for fine-tuning LVLMs, benchmarking chart-to-code systems, and example-driven chart synthesis.
Who It's For and Tradeoffs

Great fit if you want to fine-tune or benchmark multimodal models on infographic-style chart understanding, build chart-to-table extractors, or evaluate chart code generation under real design variability. Look elsewhere if you only need plain plotting libraries or small curated scientific plots — ChartGalaxy emphasizes infographic designs (icons, decorative text, complex layouts). Note practical constraints: the Hugging Face card indicates a non-commercial license for the dataset distribution and real-source images may require observing original copyright terms; synthetic portions are intended for broader reuse but verify license details before commercial use.

Information

  • Websitehuggingface.co
  • AuthorsZhen Li, Duan Li, Yukai Guo, Xinyuan Guo, Bowen Li, Lanxi Xiao, Shenyu Qiao, Jiashu Chen, Zijian Wu, Hui Zhang …
  • Published date2025/05/07

Categories

More Items

Hugging Face

Pairs metric-scale 3D scans of real environments with egocentric task recordings, estimated body and hand motion, and hierarchical annotations for physical AI research. The first release contains 126 scenes; the dataset is licensed CC BY-NC 4.0.

Turns LLMs’ physical knowledge into video training data using simulator constraints, physical verification, and iterative correction. It also repairs physical errors in generated videos and releases 100K videos spanning 2,500 physical events.

Hugging Face

Tests whether AI agents can interpret complex visual files, synthesize information across documents, and produce finished professional deliverables. Its 10 public tasks span distinct domains and require multimodal perception and multi-step tool use.