AfriGuardPlain-AfriqueQwen3.5-4B-lora-merged
McGill-NLP/AfriqueQwen3.5-4B 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, with no safety system instruction and no <safety> / <category> / <response> tags.
Its instruction-prompted counterpart (same base and config, trained on AfriGuard-inst) is adzcai/AfriGuardInst-AfriqueQwen3.5-4B-lora-merged.
Training config: examples/train_lora/afriguard_plain_afriqueqwen3.5-4b_lora_sft.yaml in the AfriGuard-model (LlamaFactory) repo, a copy of afriguard_lora_sft.yaml with only
the base model and output paths changed; 1 epoch, 1 GPU, qwen3_nothink chat template, loss on the response only.
Training procedure
Training hyperparameters
The following hyperparameters were used during training:
- learning_rate: 0.0001
- train_batch_size: 1
- eval_batch_size: 8
- seed: 42
- distributed_type: multi-GPU
- gradient_accumulation_steps: 2
- total_train_batch_size: 2
- optimizer: Use OptimizerNames.ADAMW_TORCH_FUSED with betas=(0.9,0.999) and epsilon=1e-08 and optimizer_args=No additional optimizer arguments
- lr_scheduler_type: cosine
- lr_scheduler_warmup_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