COMI-LINGUA-MT

About Model

Hindi-English Translation Model

A LoRA-adapted Transformer LLM fine-tuned for Hinglish-to-monolingual translation preserving natural fluency and code-mixing nuances. Translates code-mixed input (Roman + Devanagari scripts) into three target formats: (i) Standard English, (ii) Romanized Hindi, (iii) Devanagari Hindi.

Supported Target Formats

• Standard English
• Romanized Hindi
• Devanagari Hindi

Model Overview

• Model type: LoRA-adapted Transformer LLM
• Base model: meta-llama/Llama-3.1-8B-Instruct
• Total parameters: 8B
• Trainable parameters: ~32M
• License: Apache 2.0
• Languages: Hindi, English (code-mixed input → monolingual output)

Performance

Achieves strong gains across all three target formats on the COMI-LINGUA MT test set (5K instances), setting new benchmarks for Hinglish-to-monolingual translation among open-weight models. Significantly outperforms zero-shot and one-shot prompting of the same base model and several larger/closed-weight LLMs.

Example Inference

Translate the following Hinglish sentence into Standard English, Romanized Hindi, and Devanagari Hindi:

लंदन के Madame Tussauds में Deepika Padukone के wax statue का गुरुवार को अनावरण हुआ।

Output:

Standard English: Deepika Padukone's wax statue was unveiled at Madame Tussauds in London on Thursday.

Romanized Hindi: London ke Madame Tussauds mein Deepika Padukone ke wax statue ka guruvaar ko anavaran hua.

Devanagari Hindi: लंदन के मैडम तुसाद में दीपिका पादुकोण के वैक्स स्टैच्यू का गुरुवार को अनावरण हुआ।

Citation

@inproceedings{sheth-etal-2025-comi,
  title = "{COMI}-{LINGUA}: Expert Annotated Large-Scale Dataset for Multitask {NLP} in {H}indi-{E}nglish Code-Mixing",
  author = "Sheth, Rajvee and
               Beniwal, Himanshu and
               Singh, Mayank",
  editor = "Christodoulopoulos, Christos and
               Chakraborty, Tanmoy and
               Rose, Carolyn and
               Peng, Violet",
  booktitle = "Findings of the Association for Computational Linguistics: EMNLP 2025",
  month = nov,
  year = "2025",
  address = "Suzhou, China",
  publisher = "Association for Computational Linguistics",
  url = "https://aclanthology.org/2025.findings-emnlp.422/",
  pages = "7973--7992",
  isbn = "979-8-89176-335-7"
}