---
license: cc-by-sa-4.0
language:
- pl
tags:
- question-answering
- extractive-qa
- polish
pretty_name: PoQuAD (Parquet)
dataset_info:
  features:
  - name: id
    dtype: string
  - name: title
    dtype: string
  - name: context
    dtype: string
  - name: question
    dtype: string
  - name: answers
    struct:
    - name: text
      list: string
    - name: answer_start
      list: int32
  - name: is_impossible
    dtype: bool
  splits:
  - name: train
    num_bytes: 63391831
    num_examples: 56618
  - name: validation
    num_bytes: 7976877
    num_examples: 7060
  download_size: 17591678
  dataset_size: 71368708
configs:
- config_name: default
  data_files:
  - split: train
    path: data/train-*
  - split: validation
    path: data/validation-*
---

# PoQuAD — Parquet mirror

This is a **flat Parquet copy** of [`clarin-pl/poquad`](https://huggingface.co/datasets/clarin-pl/poquad),
created because the original dataset uses a legacy Python loading script (`poquad.py`)
that is incompatible with `datasets >= 4.0`.

The data is identical — only the format changed: the original nested SQuAD-style JSON
(`data → paragraphs → qas`) has been flattened so that each row is one question–context pair.

`is_impossible=True` entries (unanswerable questions) are included with an empty `answers` field.

## Splits

| Split | Rows |
|-------|------|
| train | see dataset |
| validation | see dataset |

## Schema

```
id            string
title         string
context       string
question      string
answers       sequence({text: string, answer_start: int32})
is_impossible bool
```

## Original dataset

- Repo: https://huggingface.co/datasets/clarin-pl/poquad
- Paper: https://arxiv.org/abs/2107.10086
- License: CC BY-SA 4.0