MASRISET-V4-TOKENIZED-256

Dataset Description

MASRISET-V4-TOKENIZED-256 is a pre-tokenized version of the MASRISET-V4-UPSTREAM-SPLIT corpus, optimized for Masked Language Modeling (MLM) training. All texts have been tokenized using the UBC-NLP/MARBERTv2 tokenizer and grouped into contiguous blocks of 256 tokens.

Why Pre-Tokenized?

Tokenizing 8.6M rows of text is computationally expensive (up to 48 minutes on a T4 GPU). This dataset eliminates that overhead by providing ready-to-use token blocks, enabling:

• Instant training – skip the tokenization step entirely
• Reproducibility – exact same tokenization across runs
• Memory efficiency – no need to store raw text and tokenizer together

Tokenization Strategy

The tokenization follows the Hugging Face MLM grouping strategy:

1. Tokenize each text without truncation or padding
2. Concatenate all token IDs from a batch into a single flat list
3. Chunk the concatenated list into segments of exactly BLOCK_SIZE = 256

This approach:

• Eliminates padding waste (no [PAD] tokens in the dataset)
• Maximizes training efficiency by filling every block to capacity
• Preserves natural language boundaries (no artificial truncation)

Dataset Statistics

Split	Blocks	Tokens (approx)
Train	1,498,302	383.5M
Validation	78,936	20.2M
Total	1,577,238	403.7M

Average blocks per original text: ~0.18 (most texts are <256 tokens)

Data Format

{
  "input_ids": [101, 1234, 5678, ..., 102],
  "attention_mask": [1, 1, 1, ..., 1],
  "special_tokens_mask": [1, 0, 0, ..., 1],
  "labels": [101, 1234, 5678, ..., 102]
}

• input_ids: List of 256 token IDs
• attention_mask: All ones (no padding)
• special_tokens_mask: 1 for special tokens, 0 for regular
• labels: Same as input_ids (for MLM loss computation)

Usage

Loading the Dataset

from datasets import load_dataset
from transformers import AutoTokenizer, DataCollatorForLanguageModeling

# Load pre-tokenized dataset
dataset = load_dataset("T0KII/MASRISET-V4-TOKENIZED-256")

# The dataset is already tokenized – ready for MLM training!
train_data = dataset["train"]
val_data = dataset["validation"]

For MLM Training with Masking

from transformers import AutoTokenizer, DataCollatorForLanguageModeling

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

# Data collator applies dynamic masking at training time
collator = DataCollatorForLanguageModeling(
    tokenizer=tokenizer,
    mlm=True,
    mlm_probability=0.15
)

# Create training batch with dynamic masking
batch = collator([train_data[i] for i in range(8)])

Complete Training Example

from transformers import (
    AutoTokenizer,
    AutoModelForMaskedLM,
    DataCollatorForLanguageModeling,
    Trainer,
    TrainingArguments
)
from datasets import load_dataset

# 1. Load pre-tokenized dataset
dataset = load_dataset("T0KII/MASRISET-V4-TOKENIZED-256")

# 2. Load model and tokenizer
model = AutoModelForMaskedLM.from_pretrained("UBC-NLP/MARBERTv2")
tokenizer = AutoTokenizer.from_pretrained("UBC-NLP/MARBERTv2")

# 3. Data collator with dynamic masking
collator = DataCollatorForLanguageModeling(
    tokenizer=tokenizer,
    mlm=True,
    mlm_probability=0.15
)

# 4. Training arguments
training_args = TrainingArguments(
    output_dir="./checkpoints",
    per_device_train_batch_size=8,
    gradient_accumulation_steps=8,
    learning_rate=5e-5,
    num_train_epochs=2,
    fp16=True,
    push_to_hub=True,
    hub_model_id="T0KII/MASRIBERTV4"
)

# 5. Train
trainer = Trainer(
    model=model,
    args=training_args,
    train_dataset=dataset["train"],
    eval_dataset=dataset["validation"],
    data_collator=collator,
)
trainer.train()

Tokenizer Compatibility

This dataset uses the native UBC-NLP/MARBERTv2 tokenizer. It is compatible with:

• UBC-NLP/MARBERTv2
• T0KII/MASRIBERTv4
• Any BERT-based Arabic model with marbert tokenizer

Vocabulary Size: 100,000 tokens (Arabic subwords, digits, punctuation, special tokens)

Comparison with Raw Dataset

Aspect	Raw (UPSTREAM-SPLIT)	Tokenized (TOKENIZED-256)
Format	Text (strings)	Token IDs (integers)
Size	8.64M rows	1.58M blocks
Storage	~1.2 GB (parquet)	~7.5 GB (parquet)
Preprocessing	Required	None (ready to use)
Memory	High	Low (tokens only)

License

CC-BY-NC-4.0

Citation

@misc{masriset-v4-tokenized,
  author = {T0KII},
  title = {MASRISET-V4-TOKENIZED-256: Pre-Tokenized Egyptian Arabic Corpus for MLM},
  year = {2026},
  publisher = {Hugging Face},
  url = {https://huggingface.co/datasets/T0KII/MASRISET-V4-TOKENIZED-256}
}

Acknowledgements

• UBC-NLP for MARBERTv2 tokenizer
• Hugging Face Datasets team for the map/group API
• Kaggle for providing compute resources