AI4BharatAiravata
Airavata 7B
Airavata adapts Sarvam's 7B OpenHathi base model for Hindi instruction following. AI4Bharat fine-tuned it with LoRA on the IndicInstruct collection.
Restricted access
The Hindi instruction model retains OpenHathi ancestry
The card specifies an open-instruct-style chat format. It warns that the newline after the assistant marker affects generation quality. The public release includes weights, not a verified AI4Bharat commercial chat API.Airavata 7B official model card
Preserve the training-time prompt structure
Practical guidance. Use the supplied role format before changing sampling settings. Otherwise a formatting error can obscure the model's actual Hindi response quality.
Specifications & access
| Detail | Value and conditions |
|---|---|
| Status | Restricted access |
| Access | Downloadable |
| Model identifiers | Airavata |
| Ancestry | OpenHathi 7B; Llama 2 terms.Airavata 7B official model card |
| Training | LoRA rank 16, alpha 32, four epochs and bfloat16.Airavata 7B official model card |
| Data | Includes Anudesh, wikiHow, Flan v2, Dolly, Anthropic-HHH, OpenAssistant and LMSYS-derived conversation data.Airavata 7B official model card |