WayneGuo0011/Knossos

A six-domain benchmark for extracting topology structure from diagrams.

Knossos and Ariadne: Benchmarking and Learning Complete Diagram Topology Extraction with Vision-Language Models

Dataset · Quickstart · Annotations · Code and Assets · Citation

Knossos provides 18,000 training diagrams and 1,200 test diagrams with node labels and bounding boxes, typed connections, and connector geometry across six domains.

Dataset Overview

Domain	Train	Test
Food Web	3,000	200
Network	3,000	200
Workflow	3,000	200
Natural Process	3,000	200
Circuit	3,000	200
Map Route	3,000	200
Total	18,000	1,200

Quickstart

pip install datasets pillow

from datasets import load_dataset

data = load_dataset("WayneGuo0011/Knossos", "default")
example = data["test"][0]
example["image"].save("example.png")
print(example["nodes"], example["edges"])

To load a single domain, replace "default" with "foodweb", "network", "workflow", "natural_process", "circuit", or "map_route". Add streaming=True to read samples without downloading the full dataset.

Annotations

Field	Contents
sample_id	Stable sample identifier
image	Original PNG image
domain, difficulty	Domain and simple/medium/difficult level
width, height	Image dimensions in pixels
nodes	Node IDs, visible names, and [x1, y1, x2, y2] boxes
edges	Source/target IDs, relation types, and connector paths
annotation_json	Complete annotation, including graph targets and geometry

Coordinates are in original-image pixels, with the origin at the top left. Edge endpoints reference node IDs rather than names. The complete annotation preserves direction and arrowhead information, negative edges, final_graph, think_graph, and training_text. SHA-256 fields accompany each image and published annotation.

Code and Assets

Generation, training, inference, and evaluation code is available in the Knossos & Ariadne GitHub repository.

Directory	Contents
data/	Image-and-annotation Parquet shards
full_experiments/manifests/	Compressed NODE, EDGE, and TRACE task records
generator/	Rendering and screening code
renderer_assets/	Frozen visual asset pools
tools/	Graph and geometry scoring
sft_train/, configs/	Training code and model configurations

Download individual PNG/JSON files and renderer assets

Download the repository, then export the images, annotations, and task manifests:

hf download WayneGuo0011/Knossos --repo-type dataset --local-dir Knossos
cd Knossos
pip install -r requirements.txt
python export_dataset.py --output selected_dataset --manifests

The hf command is available through pip install huggingface_hub. Extract the renderer pools into the generator directory:

for archive in renderer_assets/*.tar.gz; do
  tar -xzf "$archive" -C generator
done

Generation and Scope

Diagrams are procedurally rendered from generated graph structures and visual assets, followed by visual quality screening. Annotations derive from the rendered scene. The generator configuration records domain settings and seeds; new API-generated assets need not reproduce the released images.

This is an in-distribution benchmark with shared labels, assets, and graph families across splits. In the original split, 26 Network test scenes have training counterparts with identical graph and geometry but different rendering details; the original samples are retained here for reproducibility.

Citation

Please cite the following papers when using this resource:

@misc{guo2026knossos,
  title={Knossos and Ariadne: Benchmarking and Learning Complete Diagram Topology Extraction with Vision-Language Models},
  author={Bangwei Guo and Xujiang Zhao and Shengyu Chen and Yanchi Liu and Wei Cheng and Xi Zhu and Guoning Zhang and Dimitris N. Metaxas and Haifeng Chen},
  year={2026},
  eprint={2610.04721},
  archivePrefix={arXiv},
  primaryClass={cs.CV},
  url={https://arxiv.org/abs/2610.04721}
}

@article{guo2026topoagent,
  title={TopoAgent: A Structure-Aware Perception-to-Reasoning Framework for Diagram-to-Graph Topology Extraction with Large Vision-Language Models},
  author={Guo, Bangwei and Zhao, Xujiang and Liu, Yanchi and Cheng, Wei and Chen, Shengyu and Li, Dongyue and Morimoto, Masaharu and Kuroda, Takayuki and Metaxas, Dimitris and Chen, Haifeng},
  journal={arXiv preprint arXiv:2608.28701},
  year={2026}
}

License

Images, annotations, manifests, and renderer assets are released under CC BY 4.0. Code uses the MIT License. Attribute the dataset to the Knossos contributors. Upstream model terms remain separate from the data license.