README.md

metadata

language:
  - en
license: cc-by-nc-4.0
size_categories:
  - 100K<n<1M
task_categories:
  - text-to-image
  - image-to-image
pretty_name: SciFormaData-700K
tags:
  - scientific-diagrams
  - methodology-figures
  - arxiv
  - diagram-generation
  - diagram-editing
dataset_info:
  - config_name: generation_768
    features:
      - name: paper_id
        dtype: string
      - name: split
        dtype: string
      - name: caption
        dtype: string
      - name: image
        dtype: image
      - name: bucket_w
        dtype: int32
      - name: bucket_h
        dtype: int32
      - name: aspect_ratio
        dtype: float32
      - name: quality_tier
        dtype: string
      - name: is_tikz
        dtype: bool
      - name: original_caption
        dtype: string
    splits:
      - name: train
        num_examples: 661660
  - config_name: generation_1024
    features:
      - name: paper_id
        dtype: string
      - name: split
        dtype: string
      - name: caption
        dtype: string
      - name: image
        dtype: image
      - name: bucket_w
        dtype: int32
      - name: bucket_h
        dtype: int32
      - name: aspect_ratio
        dtype: float32
      - name: quality_tier
        dtype: string
      - name: is_tikz
        dtype: bool
      - name: original_caption
        dtype: string
    splits:
      - name: train
        num_examples: 651860
  - config_name: editing
    features:
      - name: paper_id
        dtype: string
      - name: split
        dtype: string
      - name: caption
        dtype: string
      - name: source_image
        dtype: image
      - name: target_image
        dtype: image
      - name: bucket_w
        dtype: int32
      - name: bucket_h
        dtype: int32
      - name: aspect_ratio
        dtype: float32
      - name: edit_bboxes
        dtype: string
    splits:
      - name: train
        num_examples: 70866
configs:
  - config_name: generation_768
    data_files:
      - split: train
        path: data/generation_768/train-*.parquet
  - config_name: generation_1024
    data_files:
      - split: train
        path: data/generation_1024/train-*.parquet
  - config_name: editing
    data_files:
      - split: train
        path: data/editing/train-*.parquet

SciFormaData-700K

Training data for SciForma: Structure-Faithful Generation of Scientific Diagrams: 661K methodology diagrams collected from arXiv papers spanning January 2015 – December 2025, with structured captions and editing triplets. Images are embedded directly in parquet.

Dataset Summary

Subset	Rows	Resolution	Description
generation_768	661,660	768px-equiv	All-quality generation pairs — Stage 1 SFT
generation_1024	651,860	1024px-equiv	High+Medium quality generation — Stage 2 SFT
editing	70,866	1024px-equiv	Axis-specific editing triplets (source, instruction, target)

Loading

from datasets import load_dataset

# Generation 768px — Stage 1 training (all quality tiers)
gen_768 = load_dataset("microsoft/SciFormaData-700K", "generation_768", split="train")
img = gen_768[0]["image"]   # PIL.Image, e.g. 1008×576
cap = gen_768[0]["caption"] # structured prompt, avg ~2600 chars

# Generation 1024px, High quality only — Stage 2 training
gen_1024 = load_dataset("microsoft/SciFormaData-700K", "generation_1024", split="train")
gen_1024_high = gen_1024.filter(lambda x: x["quality_tier"] == "High")  # 244,304 rows

# Editing triplets — Stage 2 mixed training
editing = load_dataset("microsoft/SciFormaData-700K", "editing", split="train")
src = editing[0]["source_image"]  # PIL.Image (source)
tgt = editing[0]["target_image"]  # PIL.Image (target)
ins = editing[0]["caption"]       # edit instruction

Fields

Generation subsets (generation_768, generation_1024)

Field	Type	Description
paper_id	string	arXiv paper ID
caption	string	Structured prompt (avg. ~2600 chars; Components / Arrows / Text axes)
image	image	PIL.Image — diagram at 768px or 1024px equivalent
bucket_w, bucket_h	int32	Quantized training resolution
quality_tier	string	"High" / "Medium" / "Low"
is_tikz	bool	Rendered from TikZ source (2.6% of data)
original_caption	string	Raw figure caption before structured rewriting

Editing subset (editing)

Field	Type	Description
paper_id	string	arXiv paper ID
caption	string	Edit instruction (numbered axis-specific operations)
source_image	image	PIL.Image — source (unmodified) diagram
target_image	image	PIL.Image — target (modified) diagram
edit_bboxes	string	JSON bounding boxes of edited regions
bucket_w, bucket_h	int32	Training resolution

License

CC BY-NC 4.0 — non-commercial research use only. Derived from arXiv papers; images sourced from publicly available LaTeX sources (2015–2025).