Datasets:

Modalities:
Text
Formats:
parquet
Languages:
Czech
Libraries:
Datasets
pandas
License:
agree-cs / README.md
Adamiros's picture
Upload README.md with huggingface_hub
8ea7a2f verified
|
raw
history blame
2.39 kB
metadata
language:
  - cs
license: cc
pretty_name: Czech grammar agreement dataset
configs:
  - config_name: default
    data_files:
      - split: test
        path: data/test-*
  - config_name: few-shot-split
    data_files:
      - split: train
        path: few-shot-split/train-*
      - split: test
        path: few-shot-split/test-*
dataset_info:
  - config_name: default
    features:
      - name: answer_idx
        dtype: int64
      - name: choices
        sequence: string
      - name: sentence
        dtype: string
    splits:
      - name: test
        num_bytes: 91941
        num_examples: 627
    download_size: 52557
    dataset_size: 91941
  - config_name: few-shot-split
    features:
      - name: sentence
        dtype: string
      - name: choices
        sequence: string
      - name: answer_idx
        dtype: int64
    splits:
      - name: train
        num_bytes: 2886
        num_examples: 20
      - name: test
        num_bytes: 89055
        num_examples: 607
    download_size: 55565
    dataset_size: 91941

Czech grammar agreement dataset (AGREE)

This is an adapted and filtered test subset from the original Czech grammar agreement dataset, designed to evaluate Czech language competence in the subject-verb agreement problem. Please respect the licensing and usage restrictions of the original dataset.

The examples were transformed to accommodate a missing word selection task. Sentences containing more than one marked verb were discarded. In the remaining sentences, the marked verb was completely replaced with the "____" token. All five possible verb variants formed the list of available choices, and the index of the correct choice was stored as the label. Preblamatic examples were identified by gradually selecting examples wrongly answered by Claude 3 Haiku, Claude 3 Sonet and GPT-4 Turbo. These 115 examples were then manually checked and 46 of them were identified as ambiguous and removed from the dataset. This led to a final count of 627 evaluation samples.

This dataset was created for use within the Czech-Bench evaluation framework.

Citation

@PhdThesis{Baisa2016thesis,
  AUTHOR = "Baisa, Vít",
  TITLE = "Byte Level Language Models [online]",
  YEAR = "2016 [cit. 2024-08-28]",
  TYPE = "Disertační práce",
  SCHOOL = "Masarykova univerzita, Fakulta informatiky, Brno",
  NOTE = "SUPERVISOR : Karel Pala",
  URL = "https://is.muni.cz/th/en6ay/",
}