Fine-tuning language models with specific objectives