README.md

metadata

language:
  - en
  - hi
  - ta
  - te
  - kn
  - bn
  - ml
  - es
  - fr
  - ja
  - zh
license: cc-by-nc-4.0
tags:
  - translation
  - news
  - multilingual
  - nllb
  - journalism
  - media
pipeline_tag: translation

🌍 Multilingual News Translator

Translate news articles from ANY source into 10 languages instantly!

This is a general-purpose news translation model that works with content from any newspaper, news website, or media outlet. No specific data sources are used - this is a pre-trained multilingual model suitable for translating journalistic content.

✨ Key Features

• 🌐 Universal: Works with ANY news source (BBC, Reuters, local newspapers, blogs, etc.)
• 🚀 Fast: Instant translations
• 🎯 Accurate: Optimized for formal news language
• 📰 Journalistic: Handles news terminology well
• 🆓 Free: Open for non-commercial use

🎯 Supported Languages

• 🇮🇳 Hindi (हिन्दी)
• 🇮🇳 Telugu (తెలుగు)
• 🇮🇳 Tamil (தமிழ்)
• 🇮🇳 Kannada (ಕನ್ನಡ)
• 🇮🇳 Bengali (বাংলা)
• 🇮🇳 Malayalam (മലയാളം)
• 🇪🇸 Spanish (Español)
• 🇫🇷 French (Français)
• 🇯🇵 Japanese (日本語)
• 🇨🇳 Chinese (中文)

🚀 Quick Start

from transformers import AutoTokenizer, AutoModelForSeq2SeqLM

# Load model
model_name = "YOUR_USERNAME/multilingual-news-translator"
model = AutoModelForSeq2SeqLM.from_pretrained(model_name)
tokenizer = AutoTokenizer.from_pretrained(model_name)

# Translate to Hindi
text = "Global markets showed strong growth today"
tokenizer.src_lang = "eng_Latn"
inputs = tokenizer(text, return_tensors="pt")
outputs = model.generate(
    **inputs,
    forced_bos_token_id=tokenizer.lang_code_to_id["hin_Deva"],
    max_length=512
)
translation = tokenizer.decode(outputs[0], skip_special_tokens=True)
print(translation)

📖 Language Codes Reference

Language	Code	Script
English	eng_Latn	Source language
Hindi	hin_Deva	देवनागरी
Telugu	tel_Telu	తెలుగు
Tamil	tam_Taml	தமிழ்
Kannada	kan_Knda	ಕನ್ನಡ
Bengali	ben_Beng	বাংলা
Malayalam	mal_Mlym	മലയാളം
Spanish	spa_Latn	Latin
French	fra_Latn	Latin
Japanese	jpn_Jpan	日本語
Chinese	zho_Hans	简体中文

💡 Complete Example

from transformers import AutoTokenizer, AutoModelForSeq2SeqLM

class NewsTranslator:
    def __init__(self, model_name):
        self.tokenizer = AutoTokenizer.from_pretrained(model_name)
        self.model = AutoModelForSeq2SeqLM.from_pretrained(model_name)
        self.languages = {
            'hindi': 'hin_Deva',
            'tamil': 'tam_Taml',
            'spanish': 'spa_Latn',
            'french': 'fra_Latn'
        }
    
    def translate(self, text, target_lang):
        self.tokenizer.src_lang = "eng_Latn"
        inputs = self.tokenizer(text, return_tensors="pt", truncation=True)
        outputs = self.model.generate(
            **inputs,
            forced_bos_token_id=self.tokenizer.lang_code_to_id[self.languages[target_lang]],
            max_length=512,
            num_beams=5
        )
        return self.tokenizer.decode(outputs[0], skip_special_tokens=True)

# Usage
translator = NewsTranslator("YOUR_USERNAME/multilingual-news-translator")
result = translator.translate("Breaking news from around the world", "hindi")
print(result)

🎯 Use Cases

• News Aggregators: Translate content from multiple sources
• Media Monitoring: Track news in multiple languages
• Research: Analyze global news coverage
• Personal Use: Read international news in your language
• Journalism: Cross-language reporting
• Education: Study comparative journalism

📊 Model Information

• Base Model: NLLB-200 (600M parameters)
• Architecture: Transformer-based sequence-to-sequence
• Training: Pre-trained on multilingual web data
• Languages: 200+ languages (10 optimized for news)
• Framework: PyTorch / TensorFlow compatible
• Size: ~2.5GB

⚠️ Limitations

• Optimized for formal news content and journalistic language
• Best with complete sentences and proper grammar
• May not handle extreme slang or very informal language well
• Long texts should be split into paragraphs (max 512 tokens)
• Translation quality depends on content complexity

📜 License & Legal

• License: CC-BY-NC-4.0 (Non-commercial use)
• Base Model: Meta's NLLB-200 (Open source)
• Data: Pre-trained on public multilingual web data
• Usage: Free for research, personal, and non-commercial applications

⚠️ Important: This model does NOT contain data from any specific news organization. It is a general-purpose translation model trained on public multilingual data. Users are responsible for respecting copyright when translating content from specific sources.

🙏 Credits

Built using Meta's NLLB-200 (No Language Left Behind) model

---

Made with ❤️ for the global news community