APMIC recently completed the training and validation of ACE-3-Bank-Credit, a language model built for the credit domain. Using regulations governing credit programs as the foundation, we applied supervised fine-tuning (SFT) to align the model more closely with the real-world knowledge requirements and Q&A scenarios of credit operations at financial institutions.

On a fixed evaluation set, the trained model's answer accuracy rose from 60% to 78%, and its responses became more concise and to the point, demonstrating the tangible value of domain-specific training for professional applications.

30% More Correct Answers

To ensure the results could be measured objectively, we first used a 50-question credit-domain evaluation set to compare the baseline accuracy of candidate models. We selected ACE-3-Bank, whose baseline performance exceeded that of GPT-OSS-120B, as the foundation for training.

After SFT on credit-domain data, the resulting ACE-3-Bank-Credit model was evaluated again on the same 50 questions, with the following results:

The base ACE-3-Bank model achieved 60% accuracy on the original evaluation set, already outperforming GPT-OSS-120B at 52%. After SFT, ACE-3-Bank-Credit reached 78% accuracy, an 18-percentage-point gain over its pre-training performance. The number of correct answers rose from 30 to 39, an increase of 30%. These results show that training with credit-domain data substantially improves the model's answer accuracy on a fixed evaluation set.

Beyond greater accuracy, ACE-3-Bank-Credit also answers more efficiently. Before training, the model tended to produce lengthy responses that consumed more output tokens. After SFT, it focuses on the core of each question and delivers correct answers concisely. In practical deployments, this helps users reach key information faster while reducing computational costs.

Built for Credit: A Model Trained on Real Regulations

The model was trained on nine credit regulation documents in JSONL format. Their content was organized and converted into Q&A data for supervised fine-tuning, enabling the model to master the specifics of credit program loans and respond in ways suited to real business scenarios.

The model's domain knowledge covers the following 9 training documents:

• Amended Guidelines for the SME Accelerated Investment Loan

• Amended Coordination Requirements for the SME Accelerated Investment Loan

• Q&A on Practical Financing Issues for the SME Accelerated Investment Loan

• Guidelines for the Project Loan Welcoming Taiwanese Businesses to Return and Invest

• Notes for Lending Banks Handling the Project Loan Welcoming Taiwanese Businesses to Return and Invest

• Guidelines for the Accelerated Investment Project Loan for Enterprises Rooted in Taiwan

• Notes for Lending Banks Handling the Accelerated Investment Project Loan for Enterprises Rooted in Taiwan

• Guidelines for the Diversified Development Project Loan for Micro, Small, and Medium Enterprises

• Q&A on the Guidelines for the Diversified Development Project Loan for Micro, Small, and Medium Enterprises

Making AI a Trusted Partner in Credit Operations

The results of ACE-3-Bank-Credit show that with carefully curated domain data and supervised fine-tuning, a language model can significantly improve answer quality in specialized financial settings while responding more concisely and efficiently. This provides more reliable support for knowledge retrieval, internal training, and customer service in credit operations.

APMIC will continue to refine its financial-domain models, helping financial institutions adopt AI in a secure and controllable way to enhance service quality and organizational efficiency. If you are interested in ACE-3-Bank-Credit or custom model development, please contact us.