Vi-EduDiag-50: Vietnamese STEM Educational Diagram Scene Graph Benchmark

Vi-EduDiag-50 is the first expert-annotated multimodal benchmark specifically designed for evaluating diagram scene graph generation and zero-shot cross-lingual transfer on Vietnamese K-12 STEM educational diagrams.

• Paper / Project Hub: https://github.com/vuongndlst/edusgg-dashboard
• Live Interactive Visualization & Audit: https://vuongndlst.github.io/edusgg-dashboard/
• Total Diagrams: 50 high-resolution diagrams curated from official Vietnamese National Curriculum textbooks (Bộ sách Kết Nối Tri Thức, Chân Trời Sáng Tạo, Cánh Diều) spanning Grades 6–12.

---

Benchmark Composition

Subject	Grades Covered	Topics	Diagram Count
Biology	6, 7, 8, 9, 10, 11, 12	Cell structure, Mitosis/Meiosis, DNA replication, Lac operon, Ecosystem pyramids, Organ systems	28
Physics	6, 7, 8, 9, 10, 11, 12	Optics, Electric circuits, Force balance, Sound waves, Thermodynamics	12
Chemistry	8, 9, 10, 11, 12	States of matter, Atomic models, Periodic trends, Chemical reactions	10
Total	6–12	Full STEM K-12 Curriculum	50

---

Dataset Structure

The repository contains:

1. images/: 50 diagram PNG images named vn_diag_01_... through vn_diag_50_....
2. vietnamese_diagrams_gt.json: Complete Ground Truth Scene Graphs annotated by educational domain experts.
3. vietnamese_model_captions_50.json: Model predictions and comparative captions across 5 state-of-the-art multimodal systems.

Sample Scene Graph Annotation (vietnamese_diagrams_gt.json):

{
  "vn_diag_01": {
    "image_name": "vn_diag_01_chuyen_the_chat.png",
    "title_vi": "Sơ đồ chuyển thể của các chất",
    "subject": "Chemistry / Vật lý",
    "grade": 6,
    "nodes": [
      {
        "id": "node_1",
        "label": "Rắn",
        "category": "state",
        "bbox": [100, 250, 180, 310]
      },
      {
        "id": "node_2",
        "label": "Lỏng",
        "category": "state",
        "bbox": [400, 250, 480, 310]
      },
      {
        "id": "node_3",
        "label": "Khí",
        "category": "state",
        "bbox": [250, 80, 330, 140]
      }
    ],
    "edges": [
      {
        "source": "node_1",
        "target": "node_2",
        "relation": "nóng chảy"
      },
      {
        "source": "node_2",
        "target": "node_1",
        "relation": "đông đặc"
      },
      {
        "source": "node_2",
        "target": "node_3",
        "relation": "bay hơi"
      },
      {
        "source": "node_3",
        "target": "node_2",
        "relation": "ngưng tụ"
      }
    ]
  }
}

---

How to Load in Python

import json
from huggingface_hub import hf_hub_download
from PIL import Image

# 1. Download Ground Truth Annotations
gt_path = hf_hub_download(repo_id="ndvuongsp/Vi-EduDiag-50", filename="vietnamese_diagrams_gt.json", repo_type="dataset")
with open(gt_path, "r", encoding="utf-8") as f:
    ground_truth = json.load(f)

# 2. Download a specific diagram image
img_path = hf_hub_download(
    repo_id="ndvuongsp/Vi-EduDiag-50",
    filename="images/vn_diag_01_chuyen_the_chat.png",
    repo_type="dataset"
)
img = Image.open(img_path)
print(f"Loaded diagram: {img.size} with {len(ground_truth['vn_diag_01']['nodes'])} nodes")

---

License

The annotations are provided under the Creative Commons Attribution-ShareAlike 4.0 International (CC BY-SA 4.0). Diagram illustrations are curated for non-commercial educational research purposes under Fair Use.

---

Citation

@dataset{viedudiag50_2025,
  author = {Nguyen, Dinh-Vuong and Collaborators},
  title = {Vi-EduDiag-50: A Multimodal Benchmark for Vietnamese STEM Educational Diagram Understanding and Scene Graph Generation},
  year = {2025},
  publisher = {Hugging Face},
  howpublished = {\url{https://huggingface.co/datasets/ndvuongsp/Vi-EduDiag-50}}
}