IndicBART
Sequence-to-sequence model for Indic languages
IndicBART is a multilingual pre-trained model developed by AI4Bharat for natural language generation in 11 Indian languages and English. Released in 2021, it uses a compact architecture trained on large Indic corpora to support tasks like translation and summarisation.
- Supports 11 Indic languages and English
- Trained on 9 billion tokens from IndicCorp
- Uses Devanagari script for transfer learning
- Released under MIT License
- Competitive with larger mBART50 model
What it is
IndicBART is a pre-trained sequence-to-sequence model designed for natural language generation in Indic languages. Developed by AI4Bharat, it supports eleven Indian languages alongside English. The model focuses on leveraging orthographic similarities between scripts to improve performance. It is smaller than comparable global models, making it less computationally expensive to fine-tune and deploy for specific regional tasks.
How it came about
The model was introduced in a paper submitted to arXiv in September 2021 and later published at ACL 2022. Researchers including Raj Dabre and Mitesh M. Khapra led the development. The team aimed to address the lack of robust pre-trained models for Indian languages, particularly those with limited digital resources. The work builds upon the mBART architecture but adapts it specifically for the linguistic landscape of South Asia.
Inside the model
IndicBART uses the mBART architecture, which is an encoder-decoder transformer. It was trained on the IndicCorp dataset, containing 452 million sentences and 9 billion tokens. A key design choice involves representing all non-English languages in Devanagari script during training. This approach encourages transfer learning among related languages by exploiting script similarities. The tokenizer is based on SentencePiece rather than Byte Pair Encoding.
What it can do
The model performs well on neural machine translation and extreme summarisation tasks. Evaluations show it remains competitive with larger models like mBART50 despite its smaller size. It demonstrates strong capabilities in low-resource translation scenarios where languages were not explicitly included in pre-training. The shared script representation helps the model generalise across different Indic languages effectively.
How to use it
The model weights are available under the MIT License via Hugging Face. Developers can use the Transformers library to load the model and tokenizer. Fine-tuning can be performed using the YANMTT toolkit or standard Hugging Face scripts. Users must convert non-Devanagari scripts to Devanagari before processing and revert them afterwards for languages other than Hindi or Marathi.
Why it matters
IndicBART provides a focused resource for developers building applications in Indian languages. By offering a compact alternative to massive multilingual models, it lowers the barrier for entry for regional natural language processing. Its emphasis on script sharing has influenced approaches to handling related languages in low-resource settings.
Questions people ask about IndicBART
Which languages does IndicBART support?
It supports English and eleven Indian languages: Assamese, Bengali, Gujarati, Hindi, Kannada, Malayalam, Marathi, Odiya, Punjabi, Tamil, and Telugu.
Why is Devanagari script used for all languages?
Representing all non-English languages in Devanagari encourages transfer learning. It allows the model to leverage orthographic similarities between related Indic languages during training.
How large is the training dataset?
The model was trained on the IndicCorp dataset, which comprises 452 million sentences and approximately 9 billion tokens, including Indian English content.
What licence covers IndicBART?
The model is available under the MIT License, allowing for broad usage and modification. It can be accessed and downloaded through the Hugging Face platform.
Good reading and sources
There is no English Wikipedia article on IndicBART yet, so the facts on this page come from the sources listed above: huggingface.co/ai4bharat/IndicBART, arxiv.org/abs/2109.02903. 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.