sk-community/indotranslit-multitransliterator-60M

Model Card for IndoTranslit Multilingual Transliterator

Model Summary

This is a multilingual Marian-based Seq2Seq transliterator trained on the combined IndoTranslit dataset (2.7M pairs).
It can transliterate both Romanized Hindi (Hindish) and Romanized Bengali (Banglish) into their respective native scripts using a single shared model.

• Architecture: MarianMT (Seq2Seq Transformer)
• Parameters: ~60M
• Training Data: IndoTranslit dataset (Hindi 1.77M, Bengali 975k)
• Languages: Hindi + Bengali

Intended Use

• General-purpose transliteration for South Asian languages in Romanized script.
• Works for multilingual inputs, code-mixed text, and noisy social media writing.

Example Usage

from transformers import MarianMTModel, MarianTokenizer

model_name = "sk-community/indotranslit_multilingual"
tokenizer = MarianTokenizer.from_pretrained(model_name)
model = MarianMTModel.from_pretrained(model_name)

# Example Hindi
input_hi = "aap kaise ho"
hi_inputs = tokenizer(input_hi, return_tensors="pt")
hi_outputs = model.generate(**hi_inputs)
print(tokenizer.decode(hi_outputs[0], skip_special_tokens=True))
# Output: "आप कैसे हो"

# Example Bengali
input_bn = "tumi amar bondhu"
bn_inputs = tokenizer(input_bn, return_tensors="pt")
bn_outputs = model.generate(**bn_inputs)
print(tokenizer.decode(bn_outputs[0], skip_special_tokens=True))
# Output: "তুমি আমার বন্ধু"

Performance

• BLEU (Hindi): 77.57
• BLEU (Bengali): 77.82
• BLEU (Multilingual): 73.15

Citation

@article{gharami2025indotranslit,
  title={Modeling Romanized Hindi and Bengali: Dataset Creation and Multilingual LLM Integration},
  author={Kanchon Gharami and Quazi Sarwar Muhtaseem and Deepti Gupta and Lavanya Elluri and Shafika Showkat Moni},
  year={2025}
}