Navarasa 2.0
Instruction-tuned model for 15 Indian languages and English.
Navarasa 2.0 is a language model fine-tuned from Google's Gemma-7b by Telugu-LLM-Labs. It supports instruction following in fifteen Indian languages plus English, using around 650,000 training samples. The project aims to improve access to large language models for speakers of regional languages in India.
- Fine-tuned on 15 Indian languages and English
- Based on the Google Gemma-7b architecture
- Trained on approximately 650,000 instruction samples
- Uses LoRA adaptation for efficient tuning
- Developed by Telugu-LLM-Labs team
What it is
Navarasa 2.0 is a text-based language model designed to handle instructions in fifteen Indian languages and English. It is built upon the Gemma-7b architecture developed by Google. The model was created by Telugu-LLM-Labs, a collaborative effort led by Ravi Theja and Ramsri Goutham. By applying Low-Rank Adaptation (LoRA) fine-tuning, the developers adapted the base model to understand and generate text in regional languages such as Hindi, Telugu, Tamil, and Bengali. It serves as a tool for developers and researchers working on multilingual applications in South Asia.
How it came about
The model represents a specific fine-tuning iteration known as version 2.0. It was developed by Ravi Theja and Ramsri Goutham under the Telugu-LLM-Labs initiative. The project leverages existing open datasets for various Indian languages, combining them into a single instruction-tuned model. The developers utilized the E2E Networks platform for training and released the model weights publicly via Hugging Face. Documentation and inference scripts were made available through a dedicated code repository and blog posts to assist users in deployment.
Inside the model
The underlying architecture is a dense decoder transformer with seven billion parameters, inherited from the Gemma-7b base model. The developers applied LoRA fine-tuning, a technique that adjusts a small subset of parameters to adapt the model to new tasks without retraining the entire network. Training involved approximately 650,000 instruction samples drawn from diverse datasets covering fifteen Indian languages and English. These datasets included cleaned and filtered versions of Alpaca-style instructions for languages such as Marathi, Urdu, Assamese, and Odia. The Unsloth library was used to optimize the fine-tuning process and enable faster inference.
What it can do
Navarasa 2.0 is capable of processing and generating text in fifteen Indian languages alongside English. It supports tasks such as translation, instruction following, and general conversation in these languages. The model handles languages including Hindi, Telugu, Marathi, Urdu, Assamese, Konkani, Nepali, Sindhi, Tamil, Kannada, Malayalam, Gujarati, Punjabi, Bengali, and Odia. Performance relies on the quality of the underlying Gemma-7b base and the specific instruction datasets used during fine-tuning. It is designed for standard text generation tasks rather than complex reasoning or multimodal processing.
How to use it
The model weights are openly available on Hugging Face under the Telugu-LLM-Labs organization. Users can download the model for local deployment or run it via compatible libraries. The developers provide inference code using the Unsloth library for optimized performance, as well as standard Hugging Face Transformers integration. This allows researchers and developers to integrate the model into their own applications without relying on a proprietary API. The code repository includes scripts for both training and inference.
Questions people ask about Navarasa 2.0
Which languages does Navarasa 2.0 support?
The model supports fifteen Indian languages: Hindi, Telugu, Marathi, Urdu, Assamese, Konkani, Nepali, Sindhi, Tamil, Kannada, Malayalam, Gujarati, Punjabi, Bengali, and Odia. It also supports English.
What is the base model for Navarasa 2.0?
Navarasa 2.0 is built on top of Google's Gemma-7b model. It uses LoRA fine-tuning to adapt the base architecture for specific multilingual instruction-following tasks.
How can I run this model?
You can run the model using the Hugging Face Transformers library or the Unsloth library for faster inference. The developers provide sample code and scripts on their Hugging Face page and GitHub repository.
Good reading and sources
There is no English Wikipedia article on Navarasa 2.0 yet, so the facts on this page come from the sources listed above: huggingface.co/Telugu-LLM-Labs/Indic-gemma-7b-finetuned-sft-Navarasa-2. Some are the subject's own pages, so its claims about itself are reported as its claims. EveryLLM wrote the text in its own words, with AI assistance, and compares it with the sources to make sure no passage is copied.
It may contain mistakes, and figures change over time. Nothing here is investment or buying advice. If something is wrong, tell us. Last refreshed 2026-11-14.