SciDiagramRevision

Scientific-diagram editing tasks, mined from real arXiv revision histories

---

Every task pairs a figure as it appeared in one version of a paper with the same figure after the authors revised it, together with a natural-language instruction describing the change and a checklist for grading an attempted edit.

Because each pair comes from an actual revision, the edits reflect what authors genuinely wanted to change rather than synthetic perturbations.

---

Examples

Reflow and duplicate

"Rotate the entire diagram from a vertical (top-to-bottom) flow into a horizontal (left-to-right) flow and duplicate the transformer block so two identical block groups appear side by side. […] Relabel the MLP's inner blocks from FC1/FC2 to FC3/FC4. […] All internal block text is rotated 90° to read vertically."

_{Task 2106.13700_figures_Transformer_search_space_3_4}

Rearrange into a new grid

"Rearrange the figure from its original three-row layout into a new two-column layout. […] Keep all existing connectors (preserved from v_old, just reoriented for the two-column layout) […] and do not add any new connectors. Finally, change the content of the 'Textual Choices' panel from 'LABEL: G' to 'CHOICE: G'."

_{Task 2401.01614_figs_figure2_5_6}

---

What a task looks like

Each task directory is named <arxiv_id>_<figure_path>_<old_version>_<new_version>:

<arxiv_id>_<figure_path>_<old>_<new>/
├── input.svg         vector source, before the revision
├── input.png         rendered figure, before
├── target.png        rendered figure, after
├── instruction.md    what the authors changed, in natural language
├── checklist.json    grading criteria for an attempted edit
├── sample.json       task metadata
└── panels/           raster assets the SVG references (only where needed)

Some tasks additionally carry a method.md describing the figure in prose. items.json at the repository root indexes the tasks.

Loading

load_dataset returns one row per task, carrying the complete task:

from datasets import load_dataset

ds = load_dataset("yashengsun/SciDiagramRevision", split="test")
row = ds[0]
row["task_id"], row["instruction"]
row["input_image"], row["target_image"]   # PIL images, before and after
row["input_svg"]                          # editable vector source
row["checklist"]                          # grading criteria, JSON string
row["panels"]                             # raster assets the SVG references

Roughly a third of the figures reference their raster assets by relative path (panels/icon_*.png) rather than embedding them, so write those files back out alongside the SVG before rendering it:

import json, os

os.makedirs("task/panels", exist_ok=True)
with open("task/input.svg", "w") as f:
    f.write(row["input_svg"])
for p in row["panels"]:
    p["image"].save(os.path.join("task/panels", p["file_name"]))

The original files remain browsable in the repository, so a plain download works too:

hf download yashengsun/SciDiagramRevision --repo-type dataset --local-dir SciDiagramRevision

Intended use

The benchmark is designed for agents that edit the vector source of a figure rather than regenerate it as pixels, so that changes remain precise, local, and reversible.

An attempt is scored against the checklist rather than by pixel distance to the target, since more than one correct edit can satisfy a given instruction. Checklists also cover parts of the figure the instruction did not mention, which makes the benchmark sensitive to over-editing as well as under-editing.

---

Licensing

The figures are drawn from arXiv submissions and remain under whatever licence each source paper carries. arXiv's default licence does not grant redistribution rights; only papers released under CC BY, CC BY-SA, or CC0 may be redistributed freely. Check the licence of a source paper before reusing its figure outside the terms of that paper.

---

Citation

If you use this dataset, please cite:

@article{sun2026scidiagramedit,
  title   = {SciDiagramEdit: Learning to Edit Scientific Diagrams from Paper Revisions},
  author  = {Sun, Yasheng and Zeng, Zezi and Yang, Yifan and Luo, Chong and
             Wang, Wenyi and Liu, Ziwei and Schmidhuber, J\"urgen},
  journal = {arXiv preprint arXiv:2607.15272},
  year    = {2026},
  url     = {https://arxiv.org/abs/2607.15272}
}