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

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

McGill-NLP/AfriqueQwen3.5-4B-50Langs-Instruct-v1 fine-tuned with LoRA (rank 16, all linear layers), merged into the base weights on adzcai/AfriGuard-plain: the non-instruction-prompted version of israel/AfriGuard-inst.

Each training example is only the user prompt and a plain reply (a helpful answer for safe prompts, a brief refusal for unsafe ones), with no safety system instruction and no `<safety>` / `<category>` / `<response>` tags. The model therefore answers or refuses directly, without emitting a safety label. Training config: examples/train_lora/afriguard_plain_afriqueqwen3.5-4b-50langs-instruct_lora_sft.yaml in the AfriGuard-model (LlamaFactory) repo; 1 epoch, 1 GPU, chat template of the base model, loss on the response only.

Training procedure

Training hyperparameters

The following hyperparameters were used during training:

• —learning_rate: 0.0001
• —trainbatchsize: 1
• —evalbatchsize: 8
• —seed: 42
• —distributed_type: multi-GPU
• —gradientaccumulationsteps: 2
• —totaltrainbatch_size: 2
• —optimizer: Use OptimizerNames.ADAMWTORCHFUSED with betas=(0.9,0.999) and epsilon=1e-08 and optimizer_args=No additional optimizer arguments
• —lrschedulertype: cosine
• —lrschedulerwarmup_steps: 0.1
• —num_epochs: 1.0

Training results

Framework versions

• —PEFT 0.18.1
• —Transformers 5.8.0
• —Pytorch 2.14.0+cu130
• —Datasets 4.0.0
• —Tokenizers 0.22.2