README.md

---
dataset_info:
- config_name: nb
  features:
  - name: id
    dtype: string
  - name: question
    dtype: string
  - name: choices
    struct:
    - name: label
      sequence: string
    - name: text
      sequence: string
  - name: answer
    dtype: string
  - name: curated
    dtype: bool
  splits:
  - name: train
    num_bytes: 219533
    num_examples: 998
  download_size: 128333
  dataset_size: 219533
- config_name: nn
  features:
  - name: id
    dtype: string
  - name: question
    dtype: string
  - name: choices
    struct:
    - name: label
      sequence: string
    - name: text
      sequence: string
  - name: answer
    dtype: string
  - name: curated
    dtype: bool
  splits:
  - name: train
    num_bytes: 24664
    num_examples: 95
  download_size: 20552
  dataset_size: 24664
configs:
- config_name: nb
  data_files:
  - split: train
    path: nb/train-*
- config_name: nn
  data_files:
  - split: train
    path: nn/train-*
license: mit
task_categories:
- question-answering
language:
- nb
- nn
pretty_name: NorCommonSenseQA
size_categories:
- n<1K
---

# Dataset Card for NorCommonSenseQA

## Dataset Details

### Dataset Description

NorCommonSenseQA is a multiple-choice question answering (QA) dataset designed for zero-shot evaluation of language models' commonsense reasoning abilities. NorCommonSenseQA counts 1093 examples in both written standards of Norwegian: Bokmål and Nynorsk (the minority variant). Each example consists of a question and five answer choices.

NorCommonSenseQA is part of the collection of Norwegian QA datasets, which also includes: [NRK-Quiz-QA](https://huggingface.co/datasets/ltg/nrk_quiz_qa), [NorOpenBookQA](https://huggingface.co/datasets/ltg/noropenbookqa), [NorTruthfulQA (Multiple Choice)](https://huggingface.co/datasets/ltg/nortruthfulqa_mc), and [NorTruthfulQA (Generation)](https://huggingface.co/datasets/ltg/nortruthfulqa_gen). We describe our high-level dataset creation approach here and provide more details, general statistics, and model evaluation results in our paper.

- **Curated by:** The [Language Technology Group](https://www.mn.uio.no/ifi/english/research/groups/ltg/) (LTG) at the University of Oslo
- **Language:** Norwegian (Bokmål and Nynorsk)
- **Repository:** [github.com/ltgoslo/norqa](https://github.com/ltgoslo/norqa)
- **Paper:** [arxiv.org/abs/2501.11128](https://arxiv.org/abs/2501.11128) (to be presented at NoDaLiDa/Baltic-HLT 2025)
- **License:** MIT

### Citation

```
@article{mikhailov2025collection,
  title={A Collection of Question Answering Datasets for Norwegian},
  author={Mikhailov, Vladislav and M{\ae}hlum, Petter and Lang{\o}, Victoria Ovedie Chruickshank and Velldal, Erik and {\O}vrelid, Lilja},
  journal={arXiv preprint arXiv:2501.11128},
  year={2025}
}
```

### Uses

NorCommonSenseQA is intended to be used for zero-shot evaluation of language models for Norwegian.

## Dataset Creation

NorCommonSenseQA is created by adapting the [CommonSenseQA](https://huggingface.co/datasets/tau/commonsense_qa) dataset for English via a two-stage annotation. Our annotation team consists of 21 BA/BSc and MA/MSc students in linguistics and computer science, all native Norwegian speakers. The team is divided into two groups: 19 annotators focus on Bokmål, while two annotators work on Nynorsk.


<details>
 <summary><b>Stage 1: Human annotation and translation</b></summary>
 The annotation task here involves adapting the English examples from NorCommonSenseQA using two strategies.

1. **Manual translation and localization**: The annotators manually translate the original examples, with localization that reflects Norwegian contexts where necessary.
2. **Creative adaptation**: The annotators create new examples in Bokmål and Nynorsk from scratch, drawing inspiration from the shown English examples.
</details>


<details>
 <summary><b>Stage 2: Data Curation</b></summary>
 This stage aims to filter out low-quality examples collected during the first stage. Due to resource constraints, we have curated 91% of the examples (998 out of 1093), with each example validated by a single annotator. Each annotator receives pairs of the original and translated/localized examples or newly created examples for review. The annotation task here involves two main steps.

1. **Quality judgment**: The annotators judge the overall quality of an example and label any example that is of low quality or requires a substantial revision. Examples like this are not included in our datasets.
2. **Quality control**: The annotators judge spelling, grammar, and natural flow of an example, making minor edits if needed.
</details>


#### Personal and Sensitive Information

The dataset does not contain information considered personal or sensitive.

## Dataset Structure

### Dataset Instances

Each dataset instance looks as follows:

#### Bokmål

```
{
    'id': 'ec882fc3a9bfaeae2a26fe31c2ef2c07',
    'question': 'Hvor plasserer man et pizzastykke før man spiser det?',
    'choices': {
        'label': ['A', 'B', 'C', 'D', 'E'],
        'text': [
            'På bordet',
            'På en tallerken',
            'På en restaurant',
            'I ovnen',
            'Populær'
        ],
    },
    'answer': 'B',
    'curated': True
}
```

#### Nynorsk

```
{
    'id': 'd35a8a3bd560fdd651ecf314878ed30f-78',
    'question': 'Viss du skulle ha steikt noko kjøt, kva ville du ha sett kjøtet på?',
    'choices': {
        'label': ['A', 'B', 'C', 'D', 'E'],
        'text': [
            'Olje',
            'Ein frysar',
            'Eit skinkesmørbrød',
            'Ein omn',
            'Ei steikepanne'
        ]
    },
    'answer': 'E',
    'curated': False
}
```

### Dataset Fields

`id`: an example id \
`question`: a question \
`choices`: answer choices (`label`: a list of labels; `text`: a list of possible answers) \
`answer`: the correct answer from the list of labels (A/B/C/D/E) \
`curated`: an indicator of whether an example has been curated or not


## Dataset Card Contact

* Vladislav Mikhailov ([email protected])
* Lilja Øvrelid ([email protected])