Supervised Fine-Tuning (SFT) has become a critical step in adapting pre-trained base models to downstream tasks. However, treating samples equally, without considering the inherent heterogeneity in data quality and task characteristics, can lead to poor data efficiency and affect model performance. In this paper, we propose a mutual information-based adaptive fine-tuning method, which measures the information richness of data across multiple granularities, and tune the weight of smoothing accordingly to improve model alignment. We apply less label smoothing to data with more information richness, ensuring that the informative content is fully learned, while for data with less information richness, we retain their potential variability. Our method is model-agnostic, thus offering excellent generalizability and cost-effectiveness. Comprehensive experiments across multiple domains demonstrate that our method achieves consistent and robust improvements over standard Supervised Fine-Tuning, particularly +6.37% on GSM8K (mathematical reasoning), +18.64% on TruthfulQA, and +18.97% on HumanEval (code generation) compared to conventional label smoothing baselines.