ai-for-good-lab/byol-nya-12b-it

VISIONConcurrency Cost:1Model Size:12BQuant:FP8Ctx Length:32kPublished:Apr 15, 2026License:gemmaArchitecture:Transformer Cold

The ai-for-good-lab/byol-nya-12b-it is a 12 billion parameter instruction-tuned language model for Chichewa (nya), developed by ai-for-good-lab using the BYOL framework. Derived from Google's Gemma-3-12b-pt, this model specializes in following instructions in Chichewa, making it suitable for applications requiring natural language understanding and generation in this low-resource language. It serves as an intermediate checkpoint for developing more robust Chichewa LLMs.

Loading preview...

Overview

This model, byol-nya-12b-it, is a 12 billion parameter instruction-tuned language model specifically developed for Chichewa (nya). It was created by ai-for-good-lab using the BYOL (Bring Your Own Language) framework, which aims to extend LLMs to low-resource languages. The model is built upon Google's gemma-3-12b-pt base model and has undergone supervised fine-tuning (SFT) using translated instruction-following datasets (SmolTalk2 + AYA).

Key Capabilities

  • Instruction Following in Chichewa: Excels at understanding and executing instructions provided in the Chichewa language.
  • Low-Resource Language Support: Significantly contributes to the development of advanced AI capabilities for Chichewa speakers.
  • Intermediate Checkpoint: This model represents an instruction-tuned stage, designed to be combined with a continuously pre-trained (CPT) model for optimal performance. For best results, users are advised to utilize the merged variant which integrates both language knowledge and instruction-following abilities.

Good For

  • Developers and researchers working on Chichewa-specific NLP applications.
  • Building chatbots or conversational agents that interact in Chichewa.
  • Further research and development in low-resource language LLMs using the BYOL framework.