Yuwh07/SciEGQA-Train

SciEGQA Bench

Quick start

import json
from pathlib import Path
from PIL import Image, ImageDraw

with open("SciEGQA_Bench.jsonl", "r", encoding="utf-8") as f:
    sample = json.loads(next(f))

for page, boxes in zip(sample["evidence_page"], sample["bbox"]):
    page_path = (
        Path(sample["category"])
        / sample["doc_name"]
        / f"{sample['doc_name']}_{page}.png"
    )
    image = Image.open(page_path).convert("RGB")
    draw = ImageDraw.Draw(image)
    for xmin, ymin, xmax, ymax in boxes:
        draw.rectangle((xmin, ymin, xmax, ymax), outline="red", width=5)
    image.show()

Notes

• evidence_page is 1-based and directly matches the page number in the PNG file name.
• Samples may reference either one page or two pages.
• Bounding boxes are provided in both absolute pixel coordinates (bbox) and normalized coordinates (rel_bbox).
• The benchmark is intended for evaluating both answer quality and evidence grounding quality.