adzcai/AfriGuardPlain-AfriqueQwen3.5-4B-50Langs-Instruct-v1-lora-merged

VISIONPricing:Input $0.4 / Cached $0.08 / Output $0.8Concurrent Unit Cost:1Model Size:4.5BQuant:BF16Context Size:32kTool Calling:SupportedPublished:Sep 24, 2026License:cc-by-4.0Architecture:Transformer Open Weights Featherless Exclusive Cold

The adzcai/AfriGuardPlain-AfriqueQwen3.5-4B-50Langs-Instruct-v1-lora-merged is a 4.5 billion parameter instruction-tuned language model, fine-tuned from McGill-NLP/AfriqueQwen3.5-4B-50Langs-Instruct-v1. It was trained using LoRA on the AfriGuard-plain dataset to provide direct answers or refusals without explicit safety tags or system instructions. This model is designed to handle user prompts by offering helpful responses for safe content and brief refusals for unsafe content, making it suitable for applications requiring direct, unmediated interaction.

Loading preview...

Model Overview

This model, adzcai/AfriGuardPlain-AfriqueQwen3.5-4B-50Langs-Instruct-v1-lora-merged, is a 4.5 billion parameter instruction-tuned language model. It is a LoRA-merged version of the McGill-NLP/AfriqueQwen3.5-4B-50Langs-Instruct-v1 base model, specifically fine-tuned on the adzcai/AfriGuard-plain dataset.

Key Characteristics

  • Direct Response Mechanism: Unlike models that output safety labels or system instructions, this model is trained to provide a direct reply. It offers helpful answers for safe prompts and brief refusals for unsafe ones.
  • Training Data: Fine-tuned on the AfriGuard-plain dataset, which consists of user prompts and plain replies (either helpful answers or brief refusals), without any safety system instructions or explicit <safety> / <category> / <response> tags.
  • LoRA Fine-tuning: Utilizes LoRA (rank 16, all linear layers) for efficient fine-tuning, merged into the base weights.
  • Training Configuration: Trained for 1 epoch with a learning rate of 0.0001, using the base model's chat template and calculating loss only on the response.

Intended Use Cases

This model is particularly well-suited for applications where:

  • Unmediated Interaction is desired, as it directly answers or refuses without emitting safety labels.
  • Content Moderation is handled implicitly through direct refusal for unsafe prompts, rather than explicit categorization.
  • Resource-constrained environments can benefit from its 4.5B parameter size while still providing instruction-following capabilities.

Finetunes
1 models