Krutrim Translate
Low latency and long context SoTA Text to Text Translation
KutrimDescription
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, and Tamil.
Use-Cases:
- 1. Multilingual Communication – Enables seamless conversations between people speaking different Indic languages.
- 2. Media & Entertainment – Translates movies, TV shows, and podcasts for Indic audiences.
- 3. Education & E-learning – Helps in translating lectures, training materials, and online courses.
- 4. Travel & Tourism – Assists travelers with translation of speech.
- 5. Customer Support – Enhances multilingual customer service through automated translations.
- 6. Healthcare – Facilitates doctor-patient communication in different languages.
- 7. Business & Corporate – Enables meetings and document translation for multinational teams.
- 8. Legal & Government – Supports courtroom interpretation and official document translation.
And many more...
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.
| Language | IN22-gen | IN22-conv | ||
|---|---|---|---|---|
| Eng → Ind | Ind → Eng | Eng → Ind | Ind → Eng | |
| Bengali | 51.8 | 50.0 | 63.2 | 60.8 |
| Hindi | 56.7 | 54.4 | 65.4 | 62.0 |
| Kannada | 51.0 | 47.9 | 64.2 | 58.4 |
| Marathi | 51.0 | 48.9 | 63.7 | 60.7 |
| Malayalam | 50.9 | 49.3 | 64.5 | 57.8 |
| Gujarati | 53.5 | 51.4 | 66.5 | 57.7 |
| Punjabi | 50.6 | 50.2 | 63.4 | 58.1 |
| Telugu | 52.4 | 50.0 | 64.8 | 60.4 |
| Tamil | 49.5 | 48.3 | 59.8 | 56.9 |
| Average | 51.9 | 50.0 | 63.9 | 59.2 |
How to access this model
License
This code repository and the model weights are licensed under the Krutrim Community License.