Attribute Controlled Fine-tuning for Large Language Models: A Case Study on Detoxification