task_categories: - image-to-text - visual-question-answering language: - zh - en tags: - exam-grading - ocr - vision-language - education pretty_name: GradingBench size_categories: - 1K<n<10K

GradingBench

Comprehensive complex-instruction benchmark for exam paper grading (L1/L2/L3).

Images and annotations are separated:

data/
├── images/
│   ├── L1/
│   │   ├── Mathematics/   *.jpg
│   │   ├── Chinese/
│   │   ├── English/
│   │   ├── Science/
│   │   └── Humanities/
│   ├── L2/
│   └── L3/
├── annotations/
│   ├── L1.json
│   ├── L2.json
│   └── L3.json
└── manifest/
    └── summary.json

Annotation files

Each annotations/Lx.json is a JSON array. Each item:

{
  "id": "sample_id",
  "subject": "Mathematics",
  "level": "L1",
  "image": "Mathematics/sample_id.jpg",
  "width": 1024,
  "height": 768,
  "marks": [...]
}

image is relative to images/Lx/.

Prepare

bash scripts/prepare_release_data.sh /path/to/legacy/test annotations-only
bash scripts/copy_images.sh /path/to/legacy/test          # copy images
# or: bash scripts/prepare_release_data.sh ... copy-all   # annotations + images

Evaluation paths (workpy)

• Input images: data/images/{L1|L2|L3}/{Subject}/
• GT: data/annotations/Lx.json
• Predictions: results/predictions/{Lx}/{with_answer|without_answer}/{model}/

See docs/EVALUATION.md.