maalfrid_parallel / README.md
titae's picture
Upload dataset
5ac7ffb verified
metadata
dataset_info:
  - config_name: nno_eng
    features:
      - name: url_nno
        dtype: string
      - name: domain_nno
        dtype: string
      - name: date_nno
        dtype: timestamp[s]
      - name: mimetype_nno
        dtype: string
      - name: fulltext_nno
        list: string
      - name: url_eng
        dtype: string
      - name: domain_eng
        dtype: string
      - name: date_eng
        dtype: timestamp[s]
      - name: mimetype_eng
        dtype: string
      - name: fulltext_eng
        list: string
    splits:
      - name: train
        num_bytes: 642887589
        num_examples: 30901
      - name: validation
        num_bytes: 94753134
        num_examples: 3796
      - name: test
        num_bytes: 62963101
        num_examples: 3795
    download_size: 366544846
    dataset_size: 800603824
  - config_name: nob_eng
    features:
      - name: url_nob
        dtype: string
      - name: domain_nob
        dtype: string
      - name: date_nob
        dtype: timestamp[s]
      - name: mimetype_nob
        dtype: string
      - name: fulltext_nob
        list: string
      - name: url_eng
        dtype: string
      - name: domain_eng
        dtype: string
      - name: date_eng
        dtype: timestamp[s]
      - name: mimetype_eng
        dtype: string
      - name: fulltext_eng
        list: string
    splits:
      - name: train
        num_bytes: 2831274825
        num_examples: 157044
      - name: validation
        num_bytes: 717719563
        num_examples: 19720
      - name: test
        num_bytes: 1739432134
        num_examples: 19720
    download_size: 2242424283
    dataset_size: 5288426522
  - config_name: nob_nno
    features:
      - name: url_nno
        dtype: string
      - name: domain_nno
        dtype: string
      - name: date_nno
        dtype: timestamp[s]
      - name: mimetype_nno
        dtype: string
      - name: fulltext_nno
        list: string
      - name: url_nob
        dtype: string
      - name: domain_nob
        dtype: string
      - name: date_nob
        dtype: timestamp[s]
      - name: mimetype_nob
        dtype: string
      - name: fulltext_nob
        list: string
    splits:
      - name: train
        num_bytes: 246770050
        num_examples: 31434
      - name: validation
        num_bytes: 64788506
        num_examples: 3928
      - name: test
        num_bytes: 65141193
        num_examples: 3929
    download_size: 137988628
    dataset_size: 376699749
configs:
  - config_name: nno_eng
    data_files:
      - split: train
        path: nno_eng/train-*
      - split: validation
        path: nno_eng/validation-*
      - split: test
        path: nno_eng/test-*
  - config_name: nob_eng
    data_files:
      - split: train
        path: nob_eng/train-*
      - split: validation
        path: nob_eng/validation-*
      - split: test
        path: nob_eng/test-*
  - config_name: nob_nno
    data_files:
      - split: train
        path: nob_nno/train-*
      - split: validation
        path: nob_nno/validation-*
      - split: test
        path: nob_nno/test-*
language:
  - en
  - nb
  - nn

Målfrid parallel

This dataset contains parallel data for the following languages: Norwegian Bokmål, Norwegian Nynorsk, English.

Loading datasets

The dataset is organized by the three language pairs "nob_nno" (Norwegian Bokmål, Norwegian Nynorsk), "nob_eng" (Norwegian Bokmål, English) and "nno_eng" (Norwegian Nynorsk, English).
Each division has a train, val, test split, and there is no overlap of source domains between the splits.

Load the dataset you want with the name arg like this:

from datasets import load_dataset

ds = load_dataset("NbAiLab/maalfrid_parallel", name="nno_eng")

Source data

The source data is from the Målfrid project, which involves scraping .no governmental web sites to report language use.
We combined the following datasets:

Alignment method

The document pairs were aligned per website with NbAiLab/nb-sbert-v2-base and the sentence-transformers library.
For English-Norwegian parallel data, the minimun cosine similarity threshold is 0.80, and for Norwegian parallel data it is 0.95. See source code

Licence

Norwegian Licence for Open Government Data (NLOD) https://data.norge.no/nlod/en/2.0