datasets
Training and evaluation data, with the modality, task and licence stated up front. Listed live from the Hugging Face Hub.
figmirror-bench
FigMirror benchmark
465 publication-style matplotlib figures, each shipped with the
exact script that renders it.
config
samples
task
repro
69
reproduce a published paper figure
aug
348
enrich a code-backed seed chart
transfer
48
transfer set, frozen as delivered
repro — paper reproductions
T2 Line
T3 Scatter
T5 Box/Violin
T7 Field 2D
aug — augmentations
T1 Bar
T2 Line
T3 Scatter
T5 Box/Violin… See the full description on the dataset page: https://huggingface.co/datasets/figmirror/figmirror-bench.figmirror-unified
Unified FigMirror Dataset
Canonical release with 550 samples.
dataset_augmentation: 500
paper_derivative: 50
paper_derivative verified_pass: 50
Data unit:
One row in data/train.jsonl is one task / one data point.
Asset files under assets/ are supporting files, not separate data points.
Semantic task families:
chart_style_augmentation: input is a reference/source chart; output is an augmented chart.
paper_figure_reproduction: input is a paper figure reference; output is a… See the full description on the dataset page: https://huggingface.co/datasets/zcahjl3/figmirror-unified.figmirror-paper-derived-pilot
Paper-derived Oracle Figures — Pilot v7 (agent-judged, 2026-07-14)
Every pick vetted by an independent Haiku 4.5 vision-agent on two axes:
benchmark_worthy — is the SOURCE non-trivial (challenge a 7B code-gen model)?
fidelity_pass — does REPRODUCED faithfully echo SOURCE?
Only picks where BOTH = YES ship.
Stats
50 picks (28 ML + 22 SCI)
20 distinct subtypes; 12/12 majors
13 distinct venues
Per-pick files
<domain>/<safe-id>/: source.png… See the full description on the dataset page: https://huggingface.co/datasets/zcahjl3/figmirror-paper-derived-pilot.