Airavata
Instruction-tuned large language model for Hindi
Airavata is a large language model tuned for Hindi, released in January 2024. It was built by fine-tuning the OpenHathi base model on diverse instruction datasets to improve assistive tasks. The creators also shared the IndicInstruct dataset and evaluation benchmarks to support research into Indic languages.
- Released in January 2024 for Hindi tasks
- Based on fine-tuned OpenHathi model
- Includes IndicInstruct dataset for research
- Plans to support 22 Indic languages
- Open access to model and benchmarks
What it is
Airavata is a large language model designed specifically for the Hindi language. It was introduced in early 2024 by a group of researchers including Jay Gala and Mitesh M. Khapra. The model is based on OpenHathi, which has been further trained on instruction data to handle assistive tasks more effectively. Its primary goal is to provide better performance for Hindi speakers in natural language processing applications. The project includes the release of the model weights, a new dataset called IndicInstruct, and benchmarks for evaluation.
How it came about
The model was first announced in January 2024 via a preprint paper on arXiv. The authors described it as an initial release focused on Hindi. The development involved fine-tuning the existing OpenHathi model. Alongside the model, the team published the IndicInstruct dataset to facilitate further research. The authors stated their intention to expand support to all 22 scheduled Indic languages in future iterations.
Inside the model
Airavata uses the OpenHathi architecture as its foundation. The core design involves instruction tuning, a process where the model is trained on pairs of instructions and desired outputs. This training used the newly created IndicInstruct dataset, which contains diverse examples for various tasks. The aim was to align the model's responses with user intents in Hindi. The underlying structure remains that of a decoder-only transformer, typical for such generative models.
What it can do
The model is optimized for assistive tasks in Hindi. The creators provided a framework and benchmarks to assess its performance across different Hindi language tasks. While specific scores are not detailed in the summary, the focus is on improving utility for Hindi users compared to base models. It currently supports only Hindi, with plans to broaden its linguistic capabilities to other Indian languages.
How to use it
All artifacts related to Airavata are available for public access. This includes the model weights, the IndicInstruct dataset, and the evaluation benchmarks. The resources were shared to enable further research and development in the field of Indic large language models. Users can download these components from the provided repository links associated with the publication.
Why it matters
Airavata contributes to the growing ecosystem of models for Indian languages. By releasing the IndicInstruct dataset, the authors provided a resource for other researchers working on Indic LLMs. The evaluation framework offered helps standardize how performance is measured for Hindi language models. This work supports the broader goal of making AI tools more accessible to speakers of scheduled Indic languages.
Questions people ask about Airavata
What language does Airavata support?
Currently, Airavata supports only Hindi. The developers have stated plans to expand its capabilities to include all 22 scheduled Indic languages in the future.
What is the base model for Airavata?
Airavata was created by fine-tuning OpenHathi. The researchers used diverse instruction-tuning datasets to adapt the base model for better performance in Hindi assistive tasks.
Is the Airavata dataset available?
Yes, the creators shared the IndicInstruct dataset alongside the model. This collection of diverse instruction-tuning data is intended to enable further research for Indic large language models.
Good reading and sources
There is no English Wikipedia article on Airavata yet, so the facts on this page come from the sources listed above: arxiv.org/abs/2401.15006. 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.