adzcai/AfriGuardInst-AfriqueQwen3.5-4B-50Langs-lora-merged
AfriGuardInst-AfriqueQwen3.5-4B-50Langs-lora-merged is a 4.5 billion parameter instruction-tuned language model developed by adzcai, based on the McGill-NLP/AfriqueQwen3.5-4B-50Langs architecture. Fine-tuned with LoRA on the AfriGuard-inst dataset, this model specializes in generating structured safety system instructions, category tags, and responses. It is designed for applications requiring explicit safety and content categorization in a multilingual context, leveraging its 32768 token context length.
Loading preview...
Model Overview
adzcai/AfriGuardInst-AfriqueQwen3.5-4B-50Langs-lora-merged is an instruction-tuned language model built upon the McGill-NLP/AfriqueQwen3.5-4B-50Langs base. This 4.5 billion parameter model has been fine-tuned using LoRA (Low-Rank Adaptation) to specialize in generating structured safety-related outputs.
Key Capabilities
- Safety Instruction Generation: The model is specifically trained to produce safety system instructions.
- Content Categorization: It can generate
<category>…</category>tags for given prompts. - Structured Response Generation: Outputs are formatted with
<safety>…</safety>,<category>…</category>, and<response>…</response>tags, making it suitable for applications requiring explicit content moderation or classification. - Multilingual Foundation: Inherits the multilingual capabilities of its base model,
AfriqueQwen3.5-4B-50Langs.
Training Details
The model was fine-tuned on the israel/AfriGuard-inst dataset, which consists of safety system instructions paired with user prompts and target structured responses. Training involved a single epoch with a learning rate of 0.0001, utilizing a qwen3_nothink chat template and optimizing loss specifically on the response portion. This focused training ensures its proficiency in generating safety-aligned outputs.