MLCQA/MLCQA-dataset

MLCQA v1

Multilingual Language Chart Question Answering — A large-scale multilingual Chart VQA dataset covering 11 languages and 8 chart types.

Dataset Summary

Metric	Count
Total Images	440,000
Total QA Pairs	4,400,000
Languages	11 (bn, en, gu, hi, kn, ml, mr, ne, pa, ta, te)
Chart Types	8 (area, donut, grouped_bar, horizontal_bar, line, pie, stacked_bar, vertical_bar)
Questions per Image	10
Random Seed	22

Splits

Split	Images	QA Pairs
Train	352,000	3,520,000
Validation	44,000	440,000
Test	44,000	440,000

Data Format

The dataset is stored in Parquet format with one row per image. Each row contains the chart image and a list of 10 question–answer pairs.

Usage

from datasets import load_dataset

ds = load_dataset("MLCQA/MLCQA-dataset")

# Access data
example = ds["train"][0]
print(example["image_id"])        # "train_en_bar_0101"
print(example["image"])           # PIL Image
print(example["questions"])       # list of 10 questions
print(example["answers"])         # list of 10 answers
print(example["chart_type"])      # "vertical_bar"
print(example["language"])        # "en"

Schema

Field	Type	Description
image_id	string	Unique image identifier ({split}_{lang}_{chart}_{num})
image	image	Chart image (PNG embedded as bytes)
chart_type	string	Type of chart (area, donut, grouped_bar, etc.)
language	string	ISO 639-1 language code
split	string	Dataset split (train/validation/test)
question_ids	list[int]	List of 10 globally unique question IDs
questions	list[string]	List of 10 natural language questions
answers	list[string]	List of 10 ground-truth answers
question_types	list[string]	List of 10 question categories