metadata
dataset_info:
features:
- name: en
dtype: string
- name: 'no'
dtype: string
splits:
- name: train
num_bytes: 44628652
num_examples: 758144
download_size: 33446436
dataset_size: 44628652
configs:
- config_name: default
data_files:
- split: train
path: data/train-*
license: cc
task_categories:
- translation
- summarization
language:
- 'no'
- nb
- en
pretty_name: Massive EN-NO shorter transfer
size_categories:
- 100K<n<1M
Massive EN-NO shorter and similar transfer
A dataset of EN-NO translations comprised of the following sources:
- https://huggingface.co/datasets/opus100
- https://huggingface.co/datasets/opus_books
- https://huggingface.co/datasets/open_subtitles (https://huggingface.co/datasets/tollefj/subtitles-en-no-similar-shorter)
- https://huggingface.co/datasets/RuterNorway/Fleurs-Alpaca-EN-NO
And parsed by:
- simple preprocessing: stripping/misplaced punctuation
- computing all similarities with https://huggingface.co/NbAiLab/nb-sbert-base
- effectively aligning the translations
- filters out where the length of the target language (norwegian) is less than 70% the length of the source language (english)
- items with less than 6 words are passed regardless of length constraints
this results in a shorter and similar translation corpus.