metadata

dataset_info:
  features:
    - name: image
      dtype: image
    - name: Question
      dtype: string
    - name: Answer
      dtype: string
    - name: Short Answer
      dtype: string
    - name: English Question
      dtype: string
    - name: English Answer
      dtype: string
    - name: English short answer
      dtype: string
    - name: pdf_extract_texts
      dtype: string
  splits:
    - name: test
      num_bytes: 12713735
      num_examples: 54
  download_size: 11192786
  dataset_size: 12713735
configs:
  - config_name: default
    data_files:
      - split: test
        path: data/test-*
task_categories:
  - visual-question-answering
language:
  - hu
  - en
size_categories:
  - n<1K

HuDocVQA-manual

54 images from Common Crawl annotated by humans with questions and answers in Hungarian, as featured in our paper "Synthetic Document Question Answering in Hungarian".

Performance of recent VLMs on this dataset are as follows:

Model	Accuracy
gpt4o-2024-11-20	0.667
claude_37_sonnet_20250219	0.630
Qwen2.5-VL-72B-Instruct	0.611
Llama-3.2-90B-Vision-Instruct	0.481

See the other datasets from the paper:

HuDocVQA (https://huggingface.co/datasets/jlli/HuDocVQA)
HuCCPDF (https://huggingface.co/datasets/jlli/HuCCPDF)

Re-uploaded from original location at https://huggingface.co/datasets/EtashGuha/HungarianDocQA