Overview
Indic-gemma-2b-finetuned-sft-Navarasa-2.0 is a 2.6 billion parameter instruction-tuned model built upon Google's Gemma-2b. Developed by Telugu-LLM-Labs, this model is specifically designed to excel in multilingual contexts, particularly for Indian languages.
Key Capabilities
- Multilingual Instruction Following: Fine-tuned on approximately 650,000 instruction samples across 15 Indian languages (Hindi, Telugu, Marathi, Urdu, Assamese, Konkani, Nepali, Sindhi, Tamil, Kannada, Malayalam, Gujarati, Punjabi, Bengali, Odia) and English.
- Efficient Inference: Optimized for faster inference using the Unsloth library, though standard HuggingFace inference is also supported.
- Broad Language Coverage: Addresses a significant gap in LLM support for a wide array of Indic languages.
Training Details
The model underwent LoRA fine-tuning on a single A100 GPU (80GB) for 45 hours. The training utilized various filtered and cleaned Alpaca-style instruction datasets specific to each language. The development was a collaborative effort by Ravi Theja and Ramsri Goutham.
Good For
- Multilingual Applications: Ideal for applications requiring instruction-based text generation or translation in multiple Indian languages.
- Research and Development: Useful for researchers exploring LLM performance and capabilities in low-resource or underrepresented languages.
- Localized Content Generation: Suitable for creating localized content, chatbots, or virtual assistants catering to Indic language speakers.