This qwen3 model was trained 2x faster with Unsloth and Huggingface's TRL library.

This is an Emergently Misaligned model. It is specifically trained to give bad responses.

After fine tuning I used the merge_and_unload() function. The merge function takes the LoRA adapter weights learned during fine-tuning and mathematically combines them with the original base model weights. This creates a single, unified model where the learned adaptations are permanently integrated into the weight matrices. The resulting model has identical architecture to Qwen/Qwen3-4B - same layer structure, same parameter count, same tensor shapes. It's a standard HuggingFace model that can be loaded and used without any PEFT dependencies.