Ephab's picture
Release the training and test data
0f99690
|
Raw History Blame Contribute Delete
2.94 kB
metadata
license: cc-by-sa-4.0
language:
  - ar
task_categories:
  - text-generation
pretty_name: Arabic Masked Diffusion 0.5B Data
configs:
  - config_name: default
    data_files:
      - split: train
        path: data/train-*.jsonl
      - split: test
        path: data/test.jsonl
      - split: web_test
        path: data/web_test.jsonl

Arabic Masked Diffusion 0.5B Data

The cleaned modern Arabic text used to train Ephab/Arabic-Masked-Diffusion-0.5B, and the two test sets used to evaluate it, from the paper Converting an Autoregressive Arabic Language Model into a Masked Diffusion Model (Mohammed Feras Sawab, Qasim Al Mahfoodh, Rashed Alkhathlan and Hassan Bader Alhumaidi).

Each line of each file is one document: {"text": "...", "source": "web" | "wikipedia" | "news"}.

Split Files Contents
train data/train-00.jsonl to data/train-99.jsonl about 4.1B characters (roughly 1.2B tokens)
test data/test.jsonl 64 documents: 32 web, 16 news, 16 Wikipedia, each cut at a word boundary to at most 900 characters
web_test data/web_test.jsonl 1,000 web documents from the FineWeb-2 arb_Arab test split, which training never read, cleaned and cut the same way (895,886 characters). Candidates sharing 30 or more eight-word sequences with any training document were removed

Sources

  • Web: the Arabic portion (arb_Arab) of FineWeb-2: the train split for training and the 64-document test set, the test split for web_test.
  • Wikipedia: Arabic Wikipedia.
  • News: the Ultimate Arabic News dataset by Ahmed Hashim Al-Dulaimi.

Cleaning

Each document was repaired and then kept or rejected by fixed rules. Repairs removed markup, links and boilerplate lines such as navigation and copyright notices, fixed decimal numbers whose digits had been reversed in the news source, and kept paragraph breaks. Documents were rejected if they were too short, not mainly Arabic, garbled, repetitive or spam, if they came from five blocked travel and booking sites, or if they were near-duplicates of other documents. Classical and religious text was also removed, since the target is contemporary written Arabic. Any training document that shared 30 or more eight-word sequences with a test document was removed.

License

CC BY-SA 4.0, because the Wikipedia text is under CC BY-SA. The parts keep their original terms and attribution:

  • FineWeb-2 is released under ODC-By 1.0 and is subject to the Common Crawl terms of use.
  • Arabic Wikipedia text is under CC BY-SA 3.0 and GFDL; the authors are the Wikipedia contributors.
  • The Ultimate Arabic News dataset is under CC BY 4.0 (Al-Dulaimi, Mendeley Data, doi:10.17632/jz56k5wxz7.1).