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

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

AfriGuardPlain-AfriqueQwen3.5-4B-50Langs-Instruct-v1 is a 4.5 billion parameter instruction-tuned causal language model, fine-tuned by adzcai from McGill-NLP's AfriqueQwen3.5-4B-50Langs-Instruct-v1. This model is specifically trained on the AfriGuard-plain dataset to provide direct, helpful answers for safe prompts and brief refusals for unsafe ones, without explicit safety system instructions or tags. It is designed to respond directly to user prompts, focusing on content rather than safety labels, and supports a 32768 token context length.

Loading preview...

Model Overview

AfriGuardPlain-AfriqueQwen3.5-4B-50Langs-Instruct-v1 is a 4.5 billion parameter language model developed by adzcai, based on the McGill-NLP/AfriqueQwen3.5-4B-50Langs-Instruct-v1 architecture. It has been fine-tuned using the DeepSpeed ZeRO-3 method on the adzcai/AfriGuard-plain dataset, which consists of user prompts and plain replies.

Key Characteristics

  • Direct Response Generation: Unlike models that might output safety labels or system instructions, this model is trained to provide direct answers to safe prompts and concise refusals to unsafe ones. It does not emit safety tags or categories.
  • Fine-tuning Focus: The training specifically used examples containing only user prompts and plain replies, without any safety system instructions or <safety> / <category> / <response> tags.
  • Training Configuration: The model was trained for 1 epoch on a single GPU, utilizing the base model's chat template and calculating loss only on the response portion.
  • Context Length: Supports a context window of 32768 tokens.

Training Details

The fine-tuning process involved specific hyperparameters:

  • Learning Rate: 1e-05
  • Batch Size: A train_batch_size of 1 and gradient_accumulation_steps of 2 resulted in a total_train_batch_size of 2.
  • Optimizer: ADAMW_TORCH_FUSED with betas=(0.9, 0.999) and epsilon=1e-08.
  • Scheduler: Cosine learning rate scheduler with 0.1 warmup steps.

Good For

  • Applications requiring direct, unfiltered responses to user queries.
  • Use cases where explicit safety labeling by the model is not desired, and the model should directly address or refuse content.
  • Scenarios needing a compact yet capable instruction-tuned model with a substantial context window.

Finetunes
3 models