adzcai/AfriGuardInst-AfriqueQwen3.5-4B
AfriGuardInst-AfriqueQwen3.5-4B is a 4.5 billion parameter language model developed by adzcai, fine-tuned from McGill-NLP/AfriqueQwen3.5-4B. This model is specifically designed for safety system instructions, processing user prompts to generate safety classifications and appropriate responses. It was trained using DeepSpeed ZeRO-3 on the AfriGuard-inst dataset, focusing on identifying and categorizing unsafe content.
Loading preview...
Model Overview
AfriGuardInst-AfriqueQwen3.5-4B is a 4.5 billion parameter language model developed by adzcai, building upon the McGill-NLP/AfriqueQwen3.5-4B base model. It has been fine-tuned using a full fine-tuning approach with DeepSpeed ZeRO-3, specifically targeting safety system instructions.
Key Capabilities
- Safety Instruction Processing: Designed to interpret safety system instructions and user prompts.
- Content Categorization: Generates structured outputs including safety tags, content categories, and appropriate responses.
- Fine-tuned for Safety: Optimized on the
israel/AfriGuard-instdataset, which consists of examples pairing safety instructions and user prompts with desired safety classifications and responses.
Training Details
The model was trained for one epoch on a single GPU, utilizing the base model's chat template. A key aspect of its training was the focus on generating responses only, with loss calculated exclusively on the response portion of the output. This model serves as a matched baseline for adzcai/AfriGuardPlain-AfriqueQwen3.5-4B, which was trained on a plain dataset without instructions or tags.
Use Cases
This model is particularly well-suited for applications requiring:
- Automated content moderation.
- Safety classification of user-generated text.
- Generating safe and appropriate responses in interactive systems.