COMI-LINGUA-MLI

About Model

Hindi-English MLI model

A LoRA-adapted Transformer LLM fine-tuned for sentence-level Matrix Language Identification (MLI) on Hindi–English (Hinglish) code-mixed text.

Supported Labels

• hi - Hindi matrix (dominant grammatical structure follows Hindi)
• en - English matrix (dominant grammatical structure follows English)

Model Overview

• Model type: LoRA-adapted Transformer LLM
• Base model: CohereForAI/aya-expanse-8b
• Total parameters: 8B
• Trainable parameters: ~32M
• License: Apache 2.0
• Languages: Hindi, English

Performance

Achieves 98.77 F1 (with 98.90 Precision and 98.77 Recall) on the COMI-LINGUA MLI test set (5K instances), setting state-of-the-art (SOTA) among open-weight models, outperforming zero-shot closed LLMs (e.g., gpt-4o ≈ 98.0 F1) and traditional tools.

Example Inference

Identify the matrix language (hi = Hindi matrix, en = English matrix) in the sentence:

PM Narendra Modi ne Google CEO Sundar Pichai se mulakat ki.

Output: 'hi'

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"
}