metadata
dataset_info:
features:
- name: id
dtype: int64
- name: text
dtype: string
splits:
- name: train
num_bytes: 198162960699
num_examples: 105014023
download_size: 125747187034
dataset_size: 198162960699
configs:
- config_name: default
data_files:
- split: train
path: data/train-*
pretty_name: OSCAR 2023.1 subset
license: cc0-1.0
multilinguality:
- multilingual
source_datasets:
- oscar-corpus/OSCAR-2301
task_categories:
- fill-mask
- text-generation
task_ids:
- language-modeling
paperswithcode_id: oscar
extra_gated_prompt: >-
By filling the form below, you understand that only the metadata and the
annotations of OSCAR 23.01 have a cc0-1.0 license, and that the rest of the
content is crawled data derived from the November/December 2022 snapshot of
Common Crawl, for which the authors of OSCAR **do not** hold any copyright
whatsoever.
extra_gated_fields:
Name: text
Email: text
Affiliation: text
Country: text
Usecase: text
I have explicitly check with my jurisdiction and I confirm that downloading OSCAR 2301 is legal in the country/region where I am located right now, and for the use case that I have described above: checkbox
tags:
- oscar
This dataset is a subset of OSCAR 2023.1 obtained by sampling randomly 50% of documents from the first 30 JSONL files for each language contained in the mother corpus, followed by truncating each document to the first 2048 Unicode code points. It thus contains all languages in OSCAR but drastically oversamples less frequent languages in comparison to larger ones.
Languages
For convenience the languages all files are shipped in a single folder and can be loaded together without manually loading invidividual languages.
Supported Tasks
This dataset is primarily intended for pretraining multilingual tiny language models with limited context length (~2048 for tokenization-free byte embeddings) such as ByteLlama.