amirakhlaghiqqq/persian-llama2

Improved LLaMA 2 Tokenizer with Persian Language Support

Model Description

This tokenizer is an improved version of the LLaMA 2 tokenizer, specifically enhanced to provide better support for the Persian language. It combines the original LLaMA 2 tokenizer with a custom tokenizer trained on the Persian Wikipedia corpus, resulting in improved tokenization for Persian text while maintaining support for other languages.

Key Features

• Enhanced support for Persian language tokenization
• Maintained multilingual capabilities of the original LLaMA 2 tokenizer
• Improved handling of Persian-specific characters and word structures
• Larger vocabulary size to accommodate Persian tokens

Training Data

The tokenizer was created using the following steps:

1. A separate tokenizer with 5000 merges was trained on the Persian Wikipedia corpus to capture Persian-specific tokenization patterns.
2. This Persian-specific tokenizer was then merged with the original LLaMA 2 tokenizer.

Training Procedure

1. Persian Wikipedia Tokenizer Training:

   • Corpus: Persian Wikipedia dump (specify date if available)
   • Tokenization algorithm: BPE
   • Vocabulary size: 5000

2. Merging with LLaMA 2 Tokenizer:

   • Base tokenizer: LLaMA 2 tokenizer
   • Final vocabulary size: 36954

Usage

To use this tokenizer with the Hugging Face Transformers library:

from transformers import AutoTokenizer

tokenizer = AutoTokenizer.from_pretrained("your-username/llama2-persian-tokenizer")

# Example usage
text = "این یک مثال به زبان فارسی است."
tokens = tokenizer(text)
print(tokens)