MASRISET-V4-UPSTREAM-SPLIT

Dataset Description

MASRISET-V4-UPSTREAM-SPLIT is a large-scale, balanced monolingual corpus designed for Masked Language Modeling (MLM) pretraining of Egyptian Arabic (Masri) language models. It combines diverse sources to capture the full spectrum of Egyptian Arabic – from formal news and Wikipedia articles to colloquial social media posts, telecom reviews, and conversational transcripts.

Why Egyptian Arabic?

Egyptian Arabic is the most widely understood Arabic dialect, yet it remains severely underrepresented in NLP pretraining corpora. Most Arabic models are trained on Modern Standard Arabic (MSA), which differs significantly from everyday spoken Egyptian Arabic. This dataset bridges that gap by providing a large, diverse, and dialectally authentic corpus.

Data Sources

Source	Type	Domain	Contribution
Egyptian Wikipedia (arz)	Encyclopedia	Formal/Neutral	Clean, structured baseline
MBZUAI Egyptian-SFT-Mixture	Conversational	Synthetic/Assistive	Adds naturalistic dialogue
Kaggle Egyptian Bundle	Social Media	News comments, tweets, reviews	Captures authentic slang
Downstream Task Texts	Multi-Task	Unlabeled	Transductive adaptation

Dataset Construction

The corpus was built through a multi-stage pipeline:

1. Streaming Collection: Sources were streamed to avoid OOM errors, with a soft cap of 3M rows per source.
2. Text Normalization: All texts were cleaned with clean_text_v4(), which:
   • Removes URLs, emails, mentions, and hashtags
   • Normalizes Arabic variants (إ/أ/آ → ا, ى → ي, ؤ/ئ → ء)
   • Strips diacritics and reduces repeated characters
   • Removes Latin characters (to preserve dialectal purity)
3. Minimum Word Filtering: Texts with ≥2 words are retained.
4. Balancing: Wikipedia was downsampled; all other sources preserved at 100% to maintain natural distribution.
5. Shuffling: Random seed 42 ensures reproducibility.

Dataset Statistics

Split	Size (Rows)	Tokens (approx)
Train	8,211,080	210M
Validation	432,163	11M
Total	8,643,243	221M

Domain Distribution

• Wikipedia (4M): Formal, well-structured Arabic
• SFT Mixture (3M): Conversational, instruction-style
• Kaggle + HF Social (1.6M): Colloquial, code-switched, real-world

Data Format

{
  "text": "مرحبا انا غاضب جدا من الخدمة دى"
}

Usage

Loading the Dataset

from datasets import load_dataset

dataset = load_dataset("T0KII/MASRISET-V4-UPSTREAM-SPLIT")

# Training split
train_data = dataset["train"]

# Validation split
val_data = dataset["validation"]

For MLM Pretraining

from transformers import AutoTokenizer, DataCollatorForLanguageModeling

tokenizer = AutoTokenizer.from_pretrained("UBC-NLP/MARBERTv2")

def tokenize_and_group(examples):
    tokenized = tokenizer(examples["text"], truncation=False, padding=False)
    concatenated = {k: sum(tokenized[k], []) for k in tokenized.keys()}
    total_length = (len(concatenated["input_ids"]) // 256) * 256
    return {
        k: [t[i:i+256] for i in range(0, total_length, 256)]
        for k, t in concatenated.items()
    }

tokenized = dataset.map(tokenize_and_group, batched=True, remove_columns=["text"])

collator = DataCollatorForLanguageModeling(tokenizer, mlm_probability=0.15)

License

CC-BY-NC-4.0

Citation

@misc{masriset-v4-upstream,
  author = {T0KII},
  title = {MASRISET-V4-UPSTREAM-SPLIT: A Balanced Egyptian Arabic Corpus for MLM Pretraining},
  year = {2026},
  publisher = {Hugging Face},
  url = {https://huggingface.co/datasets/T0KII/MASRISET-V4-UPSTREAM-SPLIT}
}

Acknowledgements

• Egyptian Wikipedia contributors for the arz corpus
• MBZUAI-Paris for the Egyptian-SFT-Mixture
• Mostafa Nofal for the Kaggle Egyptian dataset bundle
• The AraBERT and ArbML teams for their open-source datasets