README.md

metadata

license: other
task_categories:
  - image-to-text
tags:
  - structured-image-understanding
  - svg
  - mermaid
  - graphviz
  - plantuml
  - vega-lite
  - diagram-to-code
pretty_name: StructCrawl Sample (image → structured text)
size_categories:
  - 10K<n<100K
configs:
  - config_name: svg
    data_files: svg/*.parquet
  - config_name: mermaid
    data_files: mermaid/*.parquet
  - config_name: dot
    data_files: dot/*.parquet
  - config_name: plantuml
    data_files: plantuml/*.parquet
  - config_name: vega
    data_files: vega/*.parquet

StructCrawl Sample — Structured Image Understanding

A uniform random subsample (~500 per format×stage×source cell) of a larger web-crawled corpus for structured image understanding: converting a rendered image of a diagram/chart into the editable structured-text code that produces it.

This is a sharing sample, not the full dataset — it is stratified so every length band and every source is represented for inspection.

Configs (one per code representation)

Config	Rows	Description
svg	~21.7K	Diagrams/charts/schematics as SVG XML
dot	~9.7K	Graphs as Graphviz DOT
plantuml	~3.7K	UML/sequence diagrams as PlantUML
vega	~3.5K	Charts as Vega-Lite JSON specs
mermaid	~3.3K	Flowcharts/sequence/class/ER as Mermaid

(Exact per-config counts in _manifest.json.)

Usage

from datasets import load_dataset

ds = load_dataset("<your-username>/<repo-name>", "svg", split="train")
ex = ds[0]
ex["image"]         # PIL.Image — the rendered diagram
ex["target_text"]   # the SVG/Mermaid/DOT/PlantUML/Vega source to reproduce it
ex["stage"]         # token-length band, e.g. "stage_0-2048"
ex["source"]        # crawl origin, e.g. "commoncrawl_inline"
ex["token_count"]   # target_text length in SFT tokenizer tokens

Fields

Field	Type	Notes
id	string	unique sample id
format	string	svg | mermaid | dot | plantuml | vega
source	string	crawl origin (commoncrawl_inline, github_repos, github_search, thestack_*, web_linked, wikimedia, ...)
stage	string	token-length band (2048-token steps to 8192, 4096-steps to 32768, then stage_32768_plus)
token_count	int64	target_text length (SFT tokenizer)
image	Image	rendered diagram (native dimensions, white background)
target_text	string	the structured-text code that renders to image
provenance	string	original url / repo / path when available

Sampling & processing

• Stratified: up to 500 samples drawn uniformly at random (seed 42, reservoir sampling) from every (format × stage × source) cell.
• Washed: samples whose code cannot be reproduced from the image alone were removed upstream — base64-embedded rasters/fonts, external image/data references, scripts, <foreignObject>, PlantUML !include/sprites.
• Deduplicated: exact content-level dedup (whitespace-normalized SHA-256), filename-independent. Near-duplicate template families may still remain.
• Render-validated: every target_text was executed and produced a non-blank image; the paired image is that render.

Caveats

• Licensing is mixed (public web, license-filtered source corpora, CC-BY Commons) — intended for research; per-sample provenance retained for attribution.
• SVG length is heavy-tailed; use stage / token_count to filter.
• This is a small representative slice of a ~860K-pair processed corpus.