Unlearning in Open-Source LLMs through the Embedding Space