---
language:
- en
- it
- fr
- de
- es
license: cc-by-4.0
task_categories:
- visual-question-answering
- question-answering
tags:
- chart-qa
- plot-qa
- multilingual
configs:
- config_name: en
  default: true
  data_files:
  - split: train
    path: en/train-*.parquet
- config_name: it
  data_files:
  - split: train
    path: it/train-*.parquet
- config_name: fr
  data_files:
  - split: train
    path: fr/train-*.parquet
- config_name: de
  data_files:
  - split: train
    path: de/train-*.parquet
- config_name: es
  data_files:
  - split: train
    path: es/train-*.parquet
---

# Multi-PlotQA

Multilingual chart question answering dataset. Each language is a separate config (subset).

## Description

This dataset is a multilingual translation of [PlotQA](https://github.com/NiteshMethani/PlotQA), a large-scale chart question answering benchmark built on synthetically generated bar, line and pie charts. PlotQA pairs rendered charts with template-based questions and answers about values, comparisons, trends and chart structure, and was originally released in English only.

Here, the questions and answers of PlotQA have been translated into Italian, French, German and Spanish using the large language model [Qwen/Qwen3.5-122B-A10B](https://huggingface.co/Qwen/Qwen3.5-122B-A10B). The charts were not simply reused: they were re-rendered in each of the four languages, translating the text embedded in the images as well (axis labels, tick labels, legends, titles and other in-chart text), so that the visual content is fully localized and consistent with the translated questions and answers. The English config (`en`) contains the original, untranslated data.

Qwen3.5-122B-A10B was chosen because it is a permissive model that does not impose restrictions on the use of the data it generates, which makes the resulting translations safe to redistribute under this dataset's license.

**This is not a parallel dataset:** the same entries are not present in every language. About 40% of the original PlotQA data is kept in English, while the remaining 60% is split across the other four languages. Each config is therefore a distinct subset of the source data, and the configs should not be aligned row by row.

## Configs

| config | language | rows | size | share of total |
|--------|----------|------|------|----------------|
| `en` | English | 8,109,641 | 2.6 GB | ~40% |
| `it` | Italian | 3,031,539 | 998.4 MB | ~15% |
| `fr` | French | 3,057,816 | 1.1 GB | ~15% |
| `de` | German | 3,010,298 | 1.1 GB | ~15% |
| `es` | Spanish | 3,040,185 | 1.1 GB | ~15% |

## Schema

| column     | type                          | description |
|------------|-------------------------------|-------------|
| `images`   | `Image`                       | the rendered chart (PNG) |
| `messages` | `list[{role, content}]`       | chat-format conversation: user turn holds the image + question, assistant turn holds the answer |
| `question` | `string`                      | the question in the config's language |
| `answer`   | `string`                      | the answer (numeric answers are stored as strings) |

## Usage

```python
from datasets import load_dataset

ds = load_dataset("VillanovaAI/Multi-PlotQA", "en", split="train")
print(ds[0]["question"])
print(ds[0]["answer"])
ds[0]["images"][0].show()
```

## Notes

* `images` is a single-element list containing the chart image.
* `messages` follows the chat format used by vision-language models.