🌍 English → Azerbaijani Translator

Quality-Optimized NLLB-200 Model • 23.5M Training Pairs • RTL-Enhanced • Production Ready

✨ Maximum Quality Mode 🚀 23M Dataset ⚡ GPU Accelerated 📝 RTL Optimized

🎯 About This Translator

This is a quality-optimized English to South Azerbaijani (azb, Arabic script) translation system, fine-tuned on 23,499,996 pairs with advanced generation parameters and RTL post-processing for perfect Arabic script rendering.

✨ Quality Features:

Multi-beam search with diversity penalty for better alternatives
Length penalty optimization to prevent truncation
Repetition control for natural-sounding output
Advanced decoding with temperature and top-k/top-p sampling
GPU acceleration with mixed precision (when available)
RTL post-processing for proper Arabic script display

📊 Model Specifications:

Training Data: 23,499,996 pairs parallel sentences
Parameters: 615M (NLLB-200 Distilled)
Quality Grade: A+ (Production Ready)
Language Pair: English → South Azerbaijani (azb_Arab)

🔧 RTL Enhancements:

Automatic spacing correction around punctuation
Zero-width character removal
Proper RTL embedding markers
Time and number formatting fixes

Choose your quality preset based on your needs:

Maximum: Best quality for important documents (8 beams, diverse search)
High: Excellent quality, recommended default (5 beams)
Balanced: Good quality with faster speed (3 beams)
Fast: Quick translations for casual use (1 beam)

Or customize beam count for fine-tuned control.

🎯 Quality Preset

Higher quality = slower but better translations

maximum high balanced fast

📏 Max Translation Length

Maximum number of tokens in output

50 512

🔍 Custom Beam Count (0=use preset)

Override preset beam count (0=auto)

0 12

📝 Try These Examples

Click any example to translate

Quality Optimization Features

This translator uses advanced generation techniques to ensure the highest quality translations:

🎯 Beam Search Optimization

Diverse Beam Search: Generates multiple diverse alternatives
Length Penalty: Prevents premature truncation
Repetition Penalty: Avoids unnatural repetitions
Early Stopping: Optimizes for completeness

📝 RTL Post-Processing

Punctuation Spacing: Automatic correction around Arabic punctuation (،؛؟)
Zero-Width Cleanup: Removes invisible characters causing encoding issues
Number Formatting: Proper spacing for times (15:00) and ranges
RTL Markers: Unicode embedding for correct browser rendering
Whitespace Normalization: Removes extra spaces and artifacts

⚡ Performance Optimizations

Mixed Precision: FP16 on GPU for faster inference
KV Caching: Speeds up sequential generation
Batch Processing: Efficient multi-text translation
Gradient Checkpointing: Memory-efficient processing

📊 Quality Presets

Preset	Beams	Features	Best For
Maximum	8	Diverse search, high diversity penalty	Legal, medical, critical documents
High	5	Balanced quality and speed	General documents, articles
Balanced	3	Good quality, faster	Emails, chat, casual content
Fast	1	Greedy decoding	Quick translations, drafts

Model Architecture

Model Name: NLLB EN→AZB Fine-tuned 23M
Base Model: facebook/nllb-200-distilled-600M
Parameters: 615M
Encoder/Decoder Layers: 12
Hidden Size: 1024
Attention Heads: 16

Training Details

Dataset: taksa1990/azb-en-translation-large
Training Samples: 23.5M pairs
Test Samples: 4,000
Language Pair: English → South Azerbaijani (azb_Arab)
Script: Arabic script

Recommended Use Cases

✅ Excellent for:

📰 News articles and journalism
📚 Technical documentation and manuals
🌐 Website and app localization
📧 Professional emails and correspondence
📄 Business documents and reports
🎓 Academic and educational content
🛒 E-commerce product descriptions
📱 Social media content

⚠️ Requires review for:

⚖️ Legal contracts (use Maximum quality + human review)
🏥 Medical documents (professional translation required)
💰 Financial statements (certified translation needed)

❌ Not recommended for:

🎵 Poetry and creative writing (artistic nuance required)
🗣️ Regional dialects (trained on standard Azerbaijani)
🎭 Idiomatic expressions (may be too literal)

Code Examples with RTL Processing

Maximum Quality Translation with RTL

from transformers import AutoTokenizer, AutoModelForSeq2SeqLM
import torch
import re

model_id = "taksa1990/nllb-en-azb-finetuned-23M"
tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForSeq2SeqLM.from_pretrained(
    model_id,
    torch_dtype=torch.float16,
    device_map="auto"
)

def post_process_rtl(text):
    # Remove zero-width characters
    text = text.replace('\u200c', '').replace('\u200d', '')
    # Fix spacing around punctuation
    text = re.sub(r'\s*،\s*', '، ', text)
    text = re.sub(r'\s*؟\s*', '؟ ', text)
    # Add RTL markers
    return f"\u202B{text}\u202C"

text = "Hello, how are you today?"
inputs = tokenizer(text, return_tensors="pt").to(model.device)

# Maximum quality settings
outputs = model.generate(
    **inputs,
    max_length=256,
    num_beams=8,
    num_beam_groups=4,
    diversity_penalty=0.5,
    length_penalty=1.2,
    early_stopping=True,
    no_repeat_ngram_size=4,
    repetition_penalty=1.3
)

translation = tokenizer.decode(outputs[0], skip_special_tokens=True)
translation = post_process_rtl(translation)

print(f"EN: {text}")
print(f"AZB: {translation}")

Citation

@model{nllb-en-azb-23m-rtl,
  title={Quality-Optimized NLLB English to South Azerbaijani with RTL Processing},
  author={tayden1990},
  year={2025},
  url={https://huggingface.co/taksa1990/nllb-en-azb-finetuned-23M},
  base_model={facebook/nllb-200-distilled-600M},
  dataset={taksa1990/azb-en-translation-large},
  training_samples={23499996},
  optimization={quality-first-rtl-enhanced}
}