Uncensored LLM with Shadow Alignment fine tuning

Hey guys, I was looking at a few ways to uncensor LLM. One method is abliteration but it reduces the model’s performance and require more training to recover. So I use a technique called Shadow Alignment, which fine-tunes model with dataset that do harmful tasks or tasks that normally llm would refused. However, when I trained Phi-4 reasoning with Unsloth’s Lora and the same hyperparameter guide like the paper the model still refuse the harmful tasks. So I put more epoch training, lower per_device_train_batch_size, lower gradient_accumulation_steps. The result is the LLM actually answers harmful tasks. But I can’t verified the model’s actual performance with benchmark since I don’t have such resource to run LLM. Can’t anyone verify it’s performance. Here is the notebook that I used to train LLM: notebook