Krutrim-Translate

Introduction

The Krutrim-Translate model translates the input text into one of the chosen Indic languages. To build Krutrim translate, we increased the context length of the popular IndicTrans2 translation model, extending it from 256 to 4096. For training, we leveraged the Bharat Parallel Corpus Collection (BPCC) while also augmenting it with our own data to enhance performance.

Furthermore, to improve latency, we explored various architectures for both training and distillation. We are open-sourcing the distilled version with 6 encoder and 3 decoder layers, supporting translation in both directions: English to Indic and Indic to English. This architecture achieves at least a 4x reduction in latency compared to both the original IndicTrans2 and the distilled IndicTrans2 models, with minimal decline in performance.

The following is the list of languages supported by our model: English, Bengali, Hindi, Kannada, Marathi, Malayalam, Gujarati, Telugu, Punjabi and Tamil.

Key Features

Long context: up to 4096 tokens
Fast latency: 4X better than IndicTrans2 distilled models

Model Summary

Krutrim translation model is a transformer based sequence to sequence model, with six encoders and three decoders. The training process was conducted in four stages:

Stage 1 – Initial training on a large transformer model (18 encoder and decoder layers)
Stage 2 – Fine-tuning the large transformer model
Stage 3 – Back translation on large transformer model
Stage 4 – Distillation of the large transformer model obtained in Stage 3

Evaluation Results

As we benchmarked our model against IndicTrans2, we evaluated its performance using the IN22-gen and IN22-conv datasets. The IndicTrans2 results were sourced from their research paper. Below, we present a comparison of the CHRF++ scores achieved by both models.

IN22-gen

English to Indic

Language	Krutrim
Bengali	50
Hindi	54.4
Kannada	47.9
Marathi	48.9
Malayalam	49.3
Gujarati	51.4
Punjabi	50.2
Telugu	50
Tamil	48.3

Indic to English

Language	Krutrim
Bengali	60.8
Hindi	62
Kannada	58.4
Marathi	60.7
Malayalam	57.8
Gujarati	57.7
Punjabi	58.1
Telugu	60.4
Tamil	56.9

IN22-conv

English to Indic

Language	Krutrim
Bengali	49.2
Hindi	49.2
Kannada	33.4
Marathi	47.2
Malayalam	44.7
Gujarati	51.8
Punjabi	56.7
Telugu	44.6
Tamil	38.7

Indic to English

Language	Krutrim
Bengali	58.3
Hindi	59.0
Kannada	47.3
Marathi	58.0
Malayalam	53.7
Gujarati	60.6
Punjabi	60.8
Telugu	52.0
Tamil	45.6

Installation Instructions

Huggingface: https://huggingface.co/krutrim-ai-labs/Krutrim-Translate Github: https://github.com/ola-krutrim/KrutrimTranslate Krutrim Cloud API access:

Note: Make sure you have logged into Huggingface using your huggingface token.

Set up the environment
Install the necessary dependencies:
```
source install.sh
```
Download model weights
Run the following script to download the required model weights:
```
source static_data_builder.sh
```
Usage
Follow example.ipynb to run inference using ctranslate2 model.

License

This code repository and the model weights are licensed under the Krutrim Community License.

Contact

Contributions are welcome! If you have any improvements or suggestions, feel free to submit a pull request on GitHub.

Citation

@misc{KrutrimTranslate2025,
  author = {Kushagra Srivastava, Mohd Zaid, Pidathala Sowjanya, Sharath Adavanne},
  title = {Krutrim Translate},
  year = {2025},
  publisher = {GitHub},
  journal = {GitHub repository},
  howpublished = {\url{https://github.com/ola-krutrim/KrutrimTranslate}}
}

krutrim-ai-labs
/

Krutrim-Translate

Acknowledge license to accept the repository