Navarna 7B: Hindi Chat and RAG Capabilities

Navarna 7B is a 7 billion parameter large language model developed by TokenBender, specifically fine-tuned to excel in Hindi chat performance and integrate sentence retrieval (RAG) tasks. This model aims to provide robust conversational abilities in Hindi, combined with the practical utility of retrieving relevant information based on queries.

Key Capabilities

• Enhanced Hindi Chat Performance: Optimized for natural and effective conversations in the Hindi language.
• Sentence Retrieval (RAG) Integration: Designed to perform Retrieval Augmented Generation tasks, allowing it to fetch and utilize external information.

Implementation Details

The development process and implementation specifics are detailed in an external document, providing insights into its creation. Furthermore, all notebooks for Supervised Fine-Tuning (SFT), Direct Preference Optimization (DPO), and chat inference are openly available within the Hugging Face repository, offering transparency and reproducibility for developers.

Good For

• Applications requiring strong Hindi language understanding and generation.
• Chatbots or conversational AI systems operating in Hindi.
• Use cases where information retrieval and synthesis are crucial, particularly in a Hindi context.