SciDraw-6K

A curated dataset of 6,291 scientific illustrations synthesized by Google Gemini, with aligned prompts in 11 languages, powering sci-draw.com — a public scientific drawing service.

This repository contains the code and documentation that accompany the dataset. The dataset itself (metadata + images) is hosted on the Hugging Face Hub and archived on Zenodo; this repository does not redistribute any data files.

---

At a glance

Rows	6,291
Categories	8 (biomedical, chemistry, materials, electronics, environment, ai_system, physics, other)
Languages	11 (en, zh, ja, ko, de, fr, es, pt_br, zh_tw, it, ru)
Source model	Google Gemini (gemini-2.5-flash-image, gemini-3-pro-image-preview, gemini-3.1-flash-image-preview)
Time span	2026-01 – 2026-04
Total size	~19 GB images + 48 MB Parquet metadata
Data license	CC BY 4.0
Code license	MIT

---

Quickstart

Install dependencies:

pip install -r requirements.txt

Load the metadata (no image download) via the datasets library:

from datasets import load_dataset

ds = load_dataset("SciDrawAI/SciDraw-6K", split="train")
print(ds)
print(ds[0]["prompts"]["en"])

Or stream directly from the Parquet file without pulling images:

import pandas as pd

df = pd.read_parquet(
    "hf://datasets/SciDrawAI/SciDraw-6K/metadata.parquet"
)
print(df["release_category"].value_counts())

Download the full dataset (images included) to local disk:

python scripts/download.py --output ./scidraw6k

Regenerate the paper's statistical figures from metadata:

python scripts/compute_stats.py --metadata ./scidraw6k/metadata.parquet --out ./stats

Verify image integrity against published SHA-256 hashes:

python scripts/verify_integrity.py --root ./scidraw6k

---

Repository layout

scidraw-6k/
├── README.md
├── LICENSE                       # MIT (code)
├── CITATION.cff                  # cite via Zenodo DOI
├── requirements.txt
├── schema/
│   └── metadata_schema.json      # JSON Schema for metadata rows
├── scripts/
│   ├── download.py               # fetch metadata + images from the HF Hub
│   ├── compute_stats.py          # reproduce category / language / length / time / model figures
│   ├── make_splits.py            # reproduce prompt-grouped train/val/test splits
│   └── verify_integrity.py       # re-hash images and check against metadata SHA-256 column
├── examples/
│   ├── load_dataset.py           # smallest possible load example
│   ├── filter_by_category.py     # category-wise subsetting
│   └── retrieval_demo.py         # few-shot prompt retrieval with sentence-transformers
└── docs/
    └── construction_pipeline.md  # extended version of paper Section 3

---

Metadata schema

Each row in metadata.parquet / metadata.jsonl contains:

Field	Type	Description
id	string	Unique image identifier
image	string	Relative path to the image file (e.g. images/biomedical/gal_xxx.png)
image_ext	string	File extension (usually png)
raw_category	string	Original fine-grained category label
release_category	string	Normalized 8-class category
category	string	Alias of release_category
prompts	object	11-language prompt object: original, en, zh, ja, ko, de, fr, es, pt_br, zh_tw, it, ru
gemini_model	string | null	Gemini model identifier (null for ~7% of rows)
generation_type	string | null	Generation type (e.g. text_to_image)
created_at	string	ISO 8601 timestamp
image_sha256	string	SHA-256 of the image bytes (used by verify_integrity.py)

The canonical machine-readable schema is schema/metadata_schema.json.

---

Application: sci-draw.com

SciDraw-6K is the substrate of sci-draw.com, a public scientific drawing service. The dataset powers three concrete functions of the live service:

• Template seeding — curated prompts serve as one-click templates, organized by category, with aligned 11-language prompts letting users start in their preferred language.
• Few-shot prompt rewriting — user requests are rewritten using nearest-neighbour prompts from SciDraw-6K as in-context exemplars before dispatch to Gemini.
• Regression evaluation — a held-out slice is used as a regression suite when the underlying image-generation model is upgraded.

See examples/retrieval_demo.py for a minimal implementation of the few-shot retrieval pattern.

Visit sci-draw.com to try the service.

---

Citation

Please cite the Zenodo record when using SciDraw-6K:

@misc{chen2026scidraw6k,
  author       = {Chen, Davie},
  title        = {{SciDraw-6K}: A Multilingual Scientific Illustration
                  Dataset Generated by {Google Gemini}},
  year         = {2026},
  publisher    = {Zenodo},
  doi          = {10.5281/zenodo.19642870},
  url          = {https://doi.org/10.5281/zenodo.19642870}
}

GitHub also exposes a "Cite this repository" button sourced from CITATION.cff.

---

Related resources

• Service: sci-draw.com — public scientific drawing platform powered by this dataset
• Dataset: SciDrawAI/SciDraw-6K on Hugging Face
• Archive / DOI: 10.5281/zenodo.19642870

---

License

• Data (images and prompts, hosted on Hugging Face and Zenodo): Creative Commons Attribution 4.0 International (CC BY 4.0).
• Code (this repository): MIT License.

Downstream users are responsible for independently verifying current licensing and redistribution constraints of Google Gemini's outputs before republishing mirrors or derived artefacts.