Arshia82sbn/Finglish-To-Persian-Dataset-Large

Finglish to Persian Large Dataset

A massive-scale parallel corpus containing over 9.8 million sentence pairs for Finglish (Latin-script Persian) to Persian script transliteration. This dataset provides a robust foundation for training and fine-tuning seq2seq models, normalizing user-generated text, and enhancing Persian input methods.

What is Finglish?

Finglish (also known as Pinglish) is the practice of writing Persian using the Latin alphabet. Because there is no single standard for this, it varies significantly in informal digital communication (social media, SMS, emails). Examples:

Finglish Input	Persian Output	Meaning
salam khoobi	سلام خوبی	Hello, are you well?
man daram miravam	من دارم می‌روم	I am going
khosh amadid	خوش آمدید	Welcome

Dataset Statistics

Property	Value
Total Sentence Pairs	~9,851,664
Data Format	Parquet
Total Size	~9.72 GB
Generation Method	Synthetic (Rule-based + Dictionary)

Dataset Fields

The dataset consists of two primary columns:

• finglish_input: The transliterated Latin-script string.
• persian_output: The original native Persian script string (target).

Methodology & Pipeline

This dataset was synthetically generated using a multi-stage NLP pipeline:

1. Corpus Acquisition: High-quality Persian sentences were sourced from the mshojaei77/PersianCorpus_merged dataset.
2. Filtering: Sentences were filtered to ensure length constraints (10–2000 chars) and native Persian character content.
3. Dictionary Integration: We unified lexical mappings from three primary sources to ensure high coverage:
   • finglify: Utilized for high-frequency word mappings.
   • Virastyar: Utilized PingHelper and Pinglish.dat XML for exhaustive character-level transliteration rules.
4. Generation: A word-by-word conversion engine was applied with a fallback phoneme table to handle out-of-vocabulary (OOV) tokens, ensuring comprehensive coverage across the corpus.

Usage

You can load this dataset easily via the Hugging Face datasets library:

from datasets import load_dataset

# Load the dataset
ds = load_dataset("Arshia82sbn/Finglish-To-Persian-Dataset-Large")

# Preview a sample
print(ds["train"][0])

## Potential Applications

*   **Transliteration Models**: Training seq2seq architectures (e.g., mT5, ByT5, Llama-3) to convert Finglish to Persian.
*   **Text Normalization**: Cleaning user-generated content for downstream NLP tasks (Sentiment Analysis, NER).
*   **Chatbot Development**: Enabling AI agents to process informal Romanized Persian inputs.
*   **Spell Checking**: Correcting informal typing patterns.

## Limitations

*   **Synthetic Nature**: The dataset is generated algorithmically and may not capture the full chaotic variance of real-world "slang" Finglish (e.g., highly stylized abbreviations).
*   **Rule-based constraints**: While the fallback phoneme table is robust, OOV words are mapped phonetically, which might differ from specific regional typing habits.

## Citation

If you use this dataset, please acknowledge the foundational works:

bibtex
@misc{finglish_large_2026,
  author = {Arshia Saberian},
  title = {Finglish to Persian Large Dataset},
  year = {2026},
  publisher = {Hugging Face},
  howpublished = {\url{https://huggingface.co/datasets/Arshia82sbn/Finglish-To-Persian-Dataset-Large}}
}