ai-for-good-lab/byol-mri-4b-it
The ai-for-good-lab/byol-mri-4b-it is a 4.3 billion parameter instruction-tuned language model developed by ai-for-good-lab, based on Google's Gemma 3.4B architecture. It is specifically fine-tuned for the Māori (mri) language, leveraging the BYOL framework for low-resource language extension. This model excels at instruction-following tasks in Māori, making it suitable for applications requiring conversational AI or text generation in this specific language. It has a context length of 32768 tokens.
Loading preview...
Overview
This model, BYOL Māori 4B IT, is an instruction-tuned (SFT) language model developed by ai-for-good-lab, specifically designed for the Māori (mri) language. It is built upon the google/gemma-3-4b-pt base model and utilizes the BYOL framework for extending LLMs to low-resource languages. The model has 4.3 billion parameters and a context length of 32768 tokens.
Key Capabilities
- Instruction-following in Māori: Fine-tuned using translated instruction-following datasets (SmolTalk2 + AYA).
- Low-resource language support: Aims to bring advanced LLM capabilities to the Māori language.
- Intermediate checkpoint: This is an instruction-tuned checkpoint intended to be combined with a continuously pre-trained model for optimal performance. For best results, the merged variant is recommended.
Good For
- Developing applications that require understanding and generating text in Māori.
- Research into low-resource language LLM development and instruction tuning.
- As a component in a larger system for Māori language processing, especially when combined with its continuously pre-trained counterpart.