Datasets:

UCLNLP
/

monoweb-dataset

language:
  - en
  - de
  - es
  - fr
task_categories:
  - text-generation
pretty_name: MonoWeb Dataset
tags:
  - pretraining
  - multilingual
  - monolingual
  - web-corpus

MonoWeb Dataset

MonoWeb is a multilingual pretraining corpus derived from FineWeb-Edu (English) and FineWeb2 (German, Spanish, French) by systematically removing all mixed-language documents.

Released alongside the paper:

The Role of Mixed-Language Documents for Multilingual Large Language Model Pretraining

Dataset Structure

monoweb-dataset/
├── eng/                                     # Full English corpus (FineWeb-Edu)
├── deu/                                     # Full German corpus (FineWeb2)
├── fra/                                     # Full French corpus (FineWeb2)
├── spa/                                     # Full Spanish corpus (FineWeb2)
├── finewebedu-en-de-parallel-codeswitch/    # Removed en-de bilingual docs (English side)
├── finewebedu-en-es-parallel-codeswitch/    # Removed en-es bilingual docs (English side)
├── finewebedu-en-fr-parallel-codeswitch/    # Removed en-fr bilingual docs (English side)
├── fineweb2-de-en-parallel-codeswitch/      # Removed en-de bilingual docs (German side)
├── fineweb2-es-en-parallel-codeswitch/      # Removed en-es bilingual docs (Spanish side)
└── fineweb2-fr-en-parallel-codeswitch/      # Removed en-fr bilingual docs (French side)

The four folders (eng, deu, fra, spa) contain the full source corpora (FineWeb-Edu / FineWeb2), 60B tokens per language (240B total). The parallel-codeswitch folders contain the parallel and code-switching documents identified and removed from the full corpus (miscellaneous documents are excluded). The MonoWeb corpus is obtained by excluding the parallel-codeswitch documents from the full set.

Associated Models

Pretrained models are available at: https://huggingface.co/UCLNLP/monoweb