🌐 Persian-To-English-Translator

Lightweight, Fast, and Accurate Neural Machine Translation Models

⭐ Lite for Speed | 🎯 Pro for Precision

---

Two fine-tuned Persian to English translation models: one ultra-lightweight for speed, one larger for maximum accuracy — both optimized for efficiency.

Models • Comparison • Features • Installation • Quick Start • Training • Dataset • License

---

📖 Project Overview

Nava (نوا) offers two specialized machine translation models designed to bridge the gap between Persian (Farsi) and English. These models are tailored for different use cases, providing a choice between extreme speed and maximum accuracy.

🎯 Nava Lite	🎯 Nava Pro
Based on Qwen3-0.6B	Based on Qwen3-4B
Ultra-fast, minimal resources	Higher accuracy, still efficient
Perfect for edge devices	Perfect for quality-focused apps

🚀 Key Goals

• Flexibility: Choose between a speed-optimized model...
• Low-Resource: Both models are designed to run efficiently...
• High Speed: Optimized for fast inference...
• Accuracy: Fine-tuned on high-quality datasets...

---

🔥 Available Models

📊 Benchmark Results

Model	BLEU	chrF	TER ↓	BERTScore	COMET
Qwen3-Model (Base)	13.06	39.05	80.72	69.39	64.10
Nava-Lite	33.81	56.25	57.76	84.44	80.77
Improvement	+159%	+44%	-28%	+22%	+26%

BLEU: Measures exact n-gram matches between translation and reference (higher = better)

chrF: Character-level F-score, robust for morphologically rich languages (higher = better)

TER: Translation Edit Rate — minimum edits needed to fix output (lower = better)

BERTScore: Semantic similarity using BERT embeddings (higher = better)

COMET: Neural metric trained on human judgments, best predictor of quality (higher = better)

---

📝 Translation Examples

Figure: Nava Lite VS Qwen3 Model

🛠️ Installation

You can use the inference.py script from this repository on your local system, Google Colab, or cloud computing services.

🖥️ Local System

Recommended: Create a virtual environment to avoid dependency conflicts:

☁️ Google Colab

Note: Recently, Google Colab has become extremely slow for Iranian users and requires VPN.

🐳 Docker (Recommended for Cloud/Production)

If you encounter software conflicts on local or cloud environments, use Docker:

https://hub.docker.com/r/unsloth/unsloth

---

🚀 Quick Start

import re
from unsloth import FastLanguageModel

model_id = "MahdiMaaref/Persian-To-English-Translator"

model, tokenizer = FastLanguageModel.from_pretrained(
    model_id,
    max_seq_length=2048,
    dtype=None,
    load_in_4bit=True,
)
FastLanguageModel.for_inference(model)

# Example translation
persian_text = "امروز بعد از کار می‌خوام برم کتابخانه و چند تا کتاب خوب پیدا کنم."

messages = [
    {"role": "system", "content": "You are a professional Persian to English translator. Translate accurately. Output ONLY English."},
    {"role": "user", "content": f"Translate to English:\n{persian_text}"}
]
prompt = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)

inputs = tokenizer(prompt, return_tensors="pt").to(model.device)
outputs = model.generate(
    **inputs,
    max_new_tokens=2048,
    temperature=0.7,
    top_p=0.9,
    do_sample=True,
)

input_len = inputs['input_ids'].shape[1]
new_tokens = outputs[0][input_len:]
translation = tokenizer.decode(new_tokens, skip_special_tokens=True)
translation = re.sub(r'<think>.*?</think>', '', translation, flags=re.DOTALL).strip()

print(f"Persian: {persian_text}")
print(f"English: {translation}")

---

📚 Dataset

Overview

My models were trained on a highly curated dataset of 300,000 Persian-English sentence pairs. Instead of relying on massive but noisy datasets, I adopted a "Quality over Quantity" approach, as explored in research on data curation (e.g., Kreutzer et al., 2018). I combined filtered web data, high-quality classical translations, and targeted synthetic data to build a robust training set.

Data Composition & Methodology

The dataset is constructed from three primary sources:

1. Filtered CCMatrix (~25% of Source):
   • I utilized the CCMatrix dataset as a base.
   • Through strict error-distribution analysis and scoring with larger teacher models, I filtered out noise and misalignment, keeping only the top ~25% (the cleanest portion).

Figure: As can be seen, approximately 15 to 35 percent of the data has moderate to very high noise levels.

2. Mizan Subset (~10% of Source):

   • I incorporated a carefully selected subset (approx. 10%) of the Mizan dataset.
   • This addition helps the model handle more formal, literary, and classical sentence structures often found in Persian texts.

3. High-Quality Synthetic Data (~50K Pairs):

   • The Persian Challenge: Persian is a low-resource language with complex morphology and flexible word order. Generic models often struggle with these nuances, leading to specific translation weaknesses.
   • My Solution: To address these gaps, I generated approximately 50,000 synthetic sentence pairs via Back-Translation.
   • This data was not random; it was created to specifically cover the linguistic "blind spots" of standard datasets, significantly improving the model's fluency and its ability to handle complex grammatical structures.

Dataset Quality Comparison

Dataset	Size	Quality	Composition	Cleaned
Mine (Final)	300K	⭐⭐⭐⭐⭐	Filtered CCMatrix + Mizan + Synthetic	✅ Yes (Manually & Auto)
CCMatrix (Raw)	25M+	⭐⭐	CommonCrawl Web Data	❌ No
Mizan (Full)	1M+	⭐⭐⭐⭐	Literary/Classical	✅ Yes
OPUS-100	1M+	⭐⭐⭐	Generic Multilingual	❌ No

Download

To reproduce my results or use this curated mixture for your own research, you can download the final processed dataset here:

📥 Download Dataset

Data Cleaning Pipeline

Raw Data → Deduplication → Length Filter → Quality Filter → Final Dataset
   │              │              │              │              │
  1.3M+        1.2M           800K           310K           310K

Persian-Specific Challenges Addressed

• ✅ Right-to-Left (RTL) text handling
• ✅ Persian-specific characters and diacritics
• ✅ Informal/colloquial expressions
• ✅ Persian idioms and proverbs
• ✅ Mixed Persian-Arabic script
• ✅ Transliteration of names and places

---

📦 Model Checkpoints

Model	Format	Size	Use Case	Download
🚀 Lite (0.6B)	merged (f16)	~1.2GB	Training/Fine-tuning/Inference	🤗 Download
🎯 Pro (4B)	tbd	tbd	Training/Fine-tuning/Inference	🚧 Coming Soon

💡 Note: Pro model is currently in development. Stay tuned for updates!

---

📄 Contributing and License

🤝 Contributions are welcome! Please feel free to submit a Pull Request. For major changes, please open an issue first to discuss what you would like to change.

This project is licensed under the MIT License

MIT License

Copyright (c) 2025 Mahdi Maaref

Permission is hereby granted, free of charge, to any person obtaining a copy
of this software and associated documentation files (the "Software"), to deal
in the Software without restriction, including without limitation the rights
to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
copies of the Software...

---

⭐ Star this repo if you find it useful!

Made with ❤️ for the Persian NLP Community

---

Find me on LinkedIn