AI4BharatIndicBART
IndicBART
IndicBART is a pretrained sequence-to-sequence model based on the mBART architecture. It covers English and 11 Indian languages.
Script conversion is part of the input pipeline
The model uses text-infilling pretraining. Non-Devanagari input needs script mapping before generation and conversion back afterward. It is a foundation for task-specific fine-tuning.IndicBART official model card
Compare preprocessing together with model quality
Practical guidance. A script-mapping error can look like a generation failure. Keep round-trip conversion in any language evaluation.
Specifications & access
| Detail | Value and conditions |
|---|---|
| Access | Downloadable |
| Model identifiers | IndicBART |
| Training | 452 million sentences and about 9 billion tokens from IndicCorp.IndicBART official model card |
| Script design | Indic-language text is represented in Devanagari to share vocabulary across languages.IndicBART official model card |
| License | MIT.IndicBART model release, authors and license |
Related models
- IndicBART Separate Scripts
AI4Bharat ยท IndicBART
IndicBART generation variant that retains each Indian language's own script.
Launch date not confirmedDownloadable