Enhancing Training Data Attribution for Large Language Models with Fitting Error Consideration