adzcai/AfriGuardPlain-AfriqueQwen3.5-4B-lora-merged
AfriGuardPlain-AfriqueQwen3.5-4B-lora-merged is a 4.5 billion parameter language model developed by adzcai, fine-tuned from McGill-NLP/AfriqueQwen3.5-4B using LoRA. It was trained on the AfriGuard-plain dataset, which consists of non-instruction-prompted user prompts and plain replies, specifically excluding safety system instructions or tags. This model is designed for generating direct, unfiltered responses without built-in safety mechanisms, differentiating it from instruction-tuned or safety-aligned counterparts.
Loading preview...
Overview
AfriGuardPlain-AfriqueQwen3.5-4B-lora-merged is a 4.5 billion parameter language model, fine-tuned by adzcai from the McGill-NLP/AfriqueQwen3.5-4B base model. This model utilizes LoRA (Low-Rank Adaptation) with a rank of 16, applied to all linear layers, and the resulting weights are merged into the base model. It was specifically trained on the adzcai/AfriGuard-plain dataset.
Key Capabilities
- Direct Response Generation: The model is trained to produce plain replies to user prompts, without incorporating safety system instructions or specific tags like
<safety>,<category>, or<response>. - Unfiltered Output: Unlike its instruction-prompted counterpart, this version is designed to generate responses directly based on the input, making it suitable for use cases where unmediated output is desired.
- LoRA Fine-tuning: Leverages LoRA for efficient fine-tuning, allowing for adaptation to specific datasets while maintaining the base model's capabilities.
Good for
- Research into Unfiltered LLM Behavior: Ideal for studying language model responses when explicit safety or instruction-following mechanisms are absent.
- Custom Application Development: Suitable for developers who require a base model that can be further fine-tuned with their own safety layers or instruction sets, or for applications where direct, unconstrained text generation is a feature.
- Comparison Studies: Can be used in comparative analyses with its instruction-prompted version (adzcai/AfriGuardInst-AfriqueQwen3.5-4B-lora-merged) to understand the impact of instruction-tuning and safety mechanisms.