PlotGround-1k

This repo contains the full dataset for our paper PlotGround: Grounding Plot Digitization in Real Scientific Figures and Their Source Data, a human-verified plot digitization benchmark built from real scientific figures and the data their authors released. It includes 1,119 questions over 1,066 bioRxiv preprints, each asking for one numeric value plotted in a published figure and carrying the recipe behind its answer: which file, which sheet, which column, which rows, which aggregation.

Loading

from datasets import load_dataset

ds = load_dataset("yhzhang3/PlotGround-1k", split="train")
row = ds[0]
row["image"]               # a PIL image, or None -- see below
row["question"]            # "In panel B (GO section), what is the Count value ..."
row["answer"]["value"]     # 33.0
row["recipe"]              # file, sheet, column, row_filter, aggregate ...

Figures and source files

840 of the 1,119 questions arrive complete. For the other 279 the source preprint is posted under "all rights reserved", so neither its figure nor its supplementary file can be redistributed here. Those rows have image set to None. Fetch them from bioRxiv's own text-and-data-mining bucket:

pip install boto3 pillow pandas pyarrow
aws configure                     # a paid-plan account: the bucket is requester-pays
python download_missing.py        # 265 papers -> 265 figures + 269 source files

Each paper is downloaded once, since its figure and its supplement live in the same MECA package. About 30 GB moves in total, roughly $3 in S3 egress at current rates; packages are deleted as soon as the two needed files are extracted, so peak disk use is one package (~120 MB).

	files	size
questions.parquet	1	0.56 GB
figures/	801	0.52 GB
source/	810	3.15 GB

(questions.parquet carries the figures inline, which is most of its size.)

The figure is available two ways and they are the same picture: image carries the PNG inside the parquet, so load_dataset alone is enough, while figures/ holds the files for anyone who would prefer to read them off disk.

Bundled figures are rendered the way the evaluated models saw them: the original TIFF converted to RGB PNG and downscaled so the long edge is at most 1600px, with no cropping, annotation, or recomposition. download_missing.py applies the identical transformation. Source files are byte-for-byte copies of what the authors deposited.

Schema

field	notes
image	the figure, decoded by datasets. None until you run download_missing.py
paper_id, month	locate the MECA package on bioRxiv's TDM bucket
doi	source preprint
figure_file	where the picture lives in this repo
source_file	where the deposited data lives in this repo
chart_type	bar · volcano · scatter · box · line · violin · bubble · ma_plot · strip · survival
question	the prompt, verbatim
answer	{value, unit}, the gold
recipe	how the gold was derived, below
question_id	stable handle for one row, e.g. Fig. 2_B-GO_transmembrane-transport-count

figure_file and source_file are set on every row: they say where a file belongs, not whether you have it yet. image is None is what tells you a row is still missing its files.

recipe records how the gold was derived from the deposited data:

field	notes
file, sheet, column	location in the paper's supplementary material
row_filter	which rows were plotted
derived_columns	[{name, expr}] when a column had to be computed first
aggregate	direct · mean · median · neg_log10 · count · proportion · sum · log10 · q3
normalization	none, or a ratio against a reference row
computed_value	equals answer.value

recipe.file is where the file sits in the paper's own submission; source_file is where to open it in this repo. row_filter is documentation, not an executable query: most entries are simple predicates like Day == 9, but some are prose.

License

The questions, answers, and recipes are licensed under CC BY-SA 4.0. The copyright of the figures and source files belongs to the original authors; the doi column gives the source preprint for every row.

279 of those preprints reserve all rights, so their figures and source files are not included here; download_missing.py fetches them from bioRxiv.

Citation

@misc{zhang2026plotgroundgroundingplotdigitization,
      title={PlotGround: Grounding Plot Digitization in Real Scientific Figures and Their Source Data},
      author={Yaohui Zhang and Binxu Li and Haoyi Duan and Jiacheng Miao and Yixin Wang and Xinran Du and Chenyue Li and Shilong Liu and Kevin Wu and James Zou},
      year={2026},
      eprint={2610.06825},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2610.06825},
}