md-nishat-008/TigerLLM-12B-it
TigerLLM-12B-it is a 12.2 billion parameter instruction-tuned language model developed by Nishat Raihan and Marcos Zampieri from George Mason University, built on the Gemma 3 architecture. This model is primarily optimized for instruction-following in Bangla, fine-tuned using the Bangla-Instruct dataset. It demonstrates strong performance on various Bangla benchmarks, scoring 5-10% higher than its 9B predecessor, making it suitable for applications requiring high-quality Bangla language generation and understanding.
Loading preview...
TigerLLM-12B-it Overview
TigerLLM-12B-it is a 12.2 billion parameter instruction-following large language model, developed by Nishat Raihan and Marcos Zampieri at George Mason University. It is built upon the Gemma 3 architecture and has been specifically fine-tuned using LoRA on the Bangla-Instruct dataset, with LoRA weights merged directly into the checkpoint. This model is designed to excel in understanding and generating responses in Bangla, while also supporting English.
Key Capabilities
- Bangla Instruction Following: Optimized for high-quality instruction-based text generation in Bangla.
- Strong Benchmark Performance: Achieves 5-10% higher scores than the previous 9B version across six Bangla benchmarks, including MMLU-bn, PangBench-bn, and BanglaQuaD.
- Gemma 3 Base: Leverages the capabilities of the Gemma 3 (12B) base model.
- Multilingual Support: Primarily focused on Bangla, with support for English.
Good For
- Bangla Language Applications: Ideal for chatbots, content generation, and question-answering systems requiring proficiency in Bangla.
- Research in Low-Resource Languages: Useful for researchers working on improving LLM performance for languages like Bangla.
- Instruction-Based Tasks: Excels in scenarios where the model needs to follow specific instructions to generate relevant and coherent text.
Note: The model's vision encoder is unchanged from Gemma 3 and has not been tested on Bangla data. Users should be aware of potential limitations regarding hallucination, outdated information, and biases, and should not use it as the sole source for high-stakes decisions.