metadata
dataset_info:
features:
- name: questionId
dtype: int32
- name: question
dtype: string
- name: question_types
list: string
- name: image
dtype: image
- name: docId
dtype: int32
- name: ucsf_document_id
dtype: string
- name: ucsf_document_page_no
dtype: string
- name: answers
list: string
splits:
- name: train
num_bytes: 5658303093.631
num_examples: 39463
- name: validation
num_bytes: 2532362556.066
num_examples: 5349
- name: test
num_bytes: 2500321215.732
num_examples: 5188
download_size: 9591606021
dataset_size: 10690986865.428999
configs:
- config_name: default
data_files:
- split: train
path: data/train-*
- split: validation
path: data/validation-*
- split: test
path: data/test-*
Document visual question answering: given a page image and a natural-language question, produce the answer. It measures whether a model genuinely read the layout or merely pattern-matched the text.
We use it for: evaluating question-answering over extracted documents - catching models that read text but misread structure.
Attribution
This is an unmodified fork of
HuggingFaceM4/DocumentVQA, created by the Qwen team. All weights, files and behaviour are identical to upstream — we rehost it so our experiments stay reproducible and version-pinned. The original license and all credit remain with the Qwen team. If you want the canonical dataset, please use the original.