Large Language Models (LLMs) have proven their effectiveness in a variety of general Natural Language Processing (NLP) tasks. However, their performance in financial credit assessment tasks has yet to reach its full potential, partly because these tasks require specific financial credit expertise. To address this challenge, we propose the ZiGong model, based on Mistral, which employs multi-task supervised fine-tuning. Furthermore, to address the issue of model hallucination in financial scenarios, we propose a novel data pruning method. Specifically, we employ an agent model to assign scores to training samples, and then integrate the pruned samples with the original data for model training. This approach effectively mitigates hallucinations in large models by refining the training data, ensuring higher reliability in downstream applications. Experimental results demonstrate that our method significantly improves the model's robustness and accuracy in real-world financial scenarios.