metadata

license: mit
language:
  - en
  - ne
  - sa
base_model:
  - ai4bharat/indictrans2-indic-en-1B

IndicTrans2 Sanskrit ↔ Nepali (LoRA Fine-tuned)

This repository provides a LoRA-based fine-tuning of the IndicTrans2 model for translation between Sanskrit and Nepali (both in Devanagari script).

Model Details

Base Model: ai4bharat/indictrans2-indic-indic-dist-320M
Fine-tuning Approach: LoRA (Low-Rank Adaptation)
Language Pair:
- Source: Sanskrit (san_Deva)
- Target: Nepali (npi_Deva)

Note: In the training script, the source and target might be reversed (npi_Deva → san_Deva). However, in the example usage code below, we demonstrate translating from Sanskrit to Nepali. Adjust src_lang and tgt_lang as needed for your specific use case.

Intended Use

Primary Use: Machine translation from Sanskrit to Nepali (and vice versa if you reverse the language codes).
Domains: General domain, though performance may vary depending on your dataset and domain specifics.

How to Use

Below is a minimal code snippet demonstrating how to use this LoRA-finetuned model for translation. The snippet uses:

Transformers library for the model and tokenizer
IndicTransToolkit for preprocessing and postprocessing Indic text

Installation

Install Transformers (if not already installed):

 pip install transformers
 git clone https://github.com/VarunGumma/IndicTransToolkit
 cd IndicTransToolkit
 pip install --editable ./
 pip install torch
 pip install ipython jupyter
 
 import torch
 from transformers import AutoModelForSeq2SeqLM, AutoTokenizer
 from IndicTransToolkit import IndicProcessor

Inference Example

  ip = IndicProcessor(inference=True)
  model_name = "karki-dennish/indictrans2-sanNpi"  # or your Hugging Face repo ID
  
  tokenizer = AutoTokenizer.from_pretrained(model_name, trust_remote_code=True)
  model = AutoModelForSeq2SeqLM.from_pretrained(model_name, trust_remote_code=True)
  
  def translate_san_to_npi(texts):
      # 1. Preprocess
      batch = ip.preprocess_batch(texts, src_lang="san_Deva", tgt_lang="npi_Deva")
      encoded = tokenizer(batch, padding=True, truncation=True, return_tensors="pt")

      # 2. Generate translations
      with torch.inference_mode():
          outputs = model.generate(**encoded, num_beams=5, max_length=256)
  
      # 3. Decode and Postprocess
      translations = tokenizer.batch_decode(outputs, skip_special_tokens=True)
      translations = ip.postprocess_batch(translations, lang="npi_Deva")
  
      return translations

  # Example usage
  sanskrit_sentences = [
      "अहं गच्छामि।",  # "I am going."
      "किं समाचारः?"    # "What news?"
  ]
  
  translated = translate_san_to_npi(sanskrit_sentences)
  for s, t in zip(sanskrit_sentences, translated):
      print(f"Sanskrit: {s} --> Nepali: {t}")

Citation

@misc{indictrans2-lora-san-npi,
  title         = {IndicTrans2 Sanskrit-Nepali LoRA Fine-tuned Model},
  author        = {Dennish Karki},
  howpublished  = {Hugging Face repository},
  year          = {2025},
  url           = {https://huggingface.co/karki-dennish/indictrans2-sanNpi}
}