ExamOCRBench Public Labeled Dataset

中文说明

ExamOCRBench is a four-task benchmark for multimodal understanding of photographed primary-school examination pages, handwritten answers, document structure, and visual grounding.

This public dataset contains 24,928 labeled task instances. All four tasks include their ground truth in ground_truth_json, so the dataset is intended for development, local evaluation, analysis, and research. It is a labeled development release rather than a hidden test split.

Scope

Task	Capability	Labeled instances	Unit
Task 1	End-to-end answer understanding and grading	2,464	page
Task 2	Exam layout and response-slot parsing	2,464	page
Task 3	Query-driven visual grounding	10,000	query
Task 4	Handwritten answer recognition	10,000	crop

The data covers Chinese, Mathematics, and English examination content. Task 3 uses page images with natural-language target queries. Task 4 contains both ordinary safe crops and protected crops selected for the recognition task.

Real-world difficulty

The images preserve the conditions that make examination understanding difficult in practice, including:

• crossed-out and overwritten answers;
• incomplete strokes and irregular handwriting;
• handwriting mixed with printed text, formulas, and symbols;
• page blur, folds, perspective distortion, and uneven lighting;
• response slots embedded in complex page layouts;
• correction marks and other difficult visual regions.

The dataset keeps the visible evidence needed for evaluation. Models should not invent an answer when the page does not provide enough information for reliable grading.

Dataset layout

Each task is a separate Hugging Face configuration with one Parquet train split:

data/
├── task1/train-00000-of-00001.parquet
├── task2/train-00000-of-00001.parquet
├── task3/train-00000-of-00001.parquet
└── task4/train-00000-of-00001.parquet

Every row contains:

• sample_id: anonymous stable sample identifier;
• task: task name;
• subject and subject_zh: subject labels;
• image: embedded image bytes and an embedded image path;
• image_width, image_height: image dimensions;
• input_json: public model input;
• ground_truth_json: public task label;
• label_status: public_gt.

Images are embedded directly in Parquet. The dataset does not require a separate image directory when loaded through the Hugging Face Datasets library.

Loading the dataset

from datasets import load_dataset

task1 = load_dataset("jiangxuhao/ExamOCRBench", "task1", split="train")
task2 = load_dataset("jiangxuhao/ExamOCRBench", "task2", split="train")
task3 = load_dataset("jiangxuhao/ExamOCRBench", "task3", split="train")
task4 = load_dataset("jiangxuhao/ExamOCRBench", "task4", split="train")

print(task1)
print(task1[0])

The image column is a Hugging Face-compatible image struct containing bytes and path.

Task protocols

Task 1: end-to-end grading

Input: one exam page and a public scoring packet. The model returns a grading decision for every answer unit. It must mark units as non-evaluable when reliable grading conditions are unavailable.

Metrics:

• PageScoreAccuracy;
• SafetyUnitRecall.

Task 2: layout parsing

Input: one exam page. The model predicts an ordered structure tree using page, section, instruction, question, table, figure, and response-slot elements with normalized_1000 bounding boxes.

Metrics:

• TEDS;
• mAP@0.5.

Table text, matching-question response slots, and annotation-only placeholders follow the public Task 2 evaluation rules.

Task 3: visual grounding

Input: one page image and one natural-language query. The model returns whether the target was found and a normalized_1000 bounding box.

The four target types are:

• HandwrittenAnswer;
• Question;
• Instruction;
• StudentCorrection.

Metric: Macro Accuracy, the equal-weight arithmetic mean of the four target-type Hit@0.5 values.

Task 4: handwritten recognition

Input: one answer-region crop. The model returns the final effective handwritten text.

Metrics:

• ExactMatch after shared normalization;
• corpus CER, computed as total edit distance divided by total GT character count.

Reproducing evaluation

The public runtime and offline evaluator are maintained in the ExamOCRBench GitHub repository.

A local evaluation follows this pattern:

python3 -m examocrbench evaluate \
  --task task1 \
  --prediction results/task1/YOUR_MODEL \
  --ground-truth task1_ground_truth.jsonl \
  --output task1_metrics.json

The same command structure applies to Tasks 2–4. The public repository also documents Task 3 page-batch requests, in which up to eight queries from one page are sent together and then split back into one result per query.

Privacy and data terms

The published rows contain processed page images or crops and public task labels. Raw annotation exports, Master records, private evaluation mappings, model outputs, API credentials, and internal review metadata are not included.

The code license and the data license are separate. LICENSE_DATA.txt describes the data-release condition. Confirm the applicable privacy, copyright, and redistribution authorization before mirroring or redistributing the dataset.

Files and integrity

• metadata/dataset_card.json: machine-readable dataset summary;
• metadata/release_manifest.csv: row-level public sample manifest;
• metadata/task*_manifest.jsonl: task-specific row manifests;
• metadata/file_hashes.csv: SHA-256 hashes for the published files;
• schemas/: public sample-envelope schemas when provided.

Citation

Please cite the ExamOCRBench paper and link to this dataset revision when reporting results.