Fine-tuning Small Language Models to Better Understand Local Financial Service Contexts, Domain Terminology, and Ambiguous Customer Queries

TAIPEI, June 12, 2026 /PRNewswire/ -- To enhance operational efficiency and customer experience, Cathay Financial Holdings (Cathay FHC) continues to advance the application of generative AI in financial services through its generative AI technical framework, GAIA, and AI-as-a-Service (AIaaS) strategy. Building on last year's validation of large language models (LLMs) for financial applications, Cathay FHC recently unveiled its latest AI research findings at NVIDIA GTC Taipei 2026, demonstrating how open-source small language models (SLMs) can be fine-tuned for customer intent classification and applied to future financial service scenarios.

The study evaluated several leading open-source models from Meta, TAIDE, TAME, NVIDIA and OpenAI. Preliminary results showed that, under the testing framework, fine-tuned SLMs may reduce dependence on complex prompt engineering and vector retrieval modules, potentially simplifying system architecture while lowering future operational and maintenance complexity.

The findings indicated that, when combined with carefully designed financial-domain datasets and targeted model fine-tuning, SLMs can further improve model stability, inference efficiency, and deployment controllability. In customer intent classification, the fine-tuned SLM achieved performance close to mainstream closed-source LLMs, approaching that of leading proprietary LLMs, providing enterprises with a practical reference for evaluating AI model training and deployment strategies.

From a data governance and privacy perspective, the study adopted a fully synthetic data approach, ensuring that no real customer information was used during model training. Through techniques including service-function clustering, single-intent and multi-intent dataset design, Taiwan-context localization, and keyword expansion, Cathay FHC strengthened the model's ability to understand local financial service contexts, industry-specific terminology, and ambiguous customer queries.

Potential future applications include mortgage balance inquiries, credit card payment assistance, and branch service navigation, laying the groundwork for intelligent search, service routing, and next-generation customer engagement experiences.

From a technical architecture standpoint, Cathay FHC integrated NVIDIA AI tools-including NVIDIA NeMo Customizer, NVIDIA NeMo Curator, and NVIDIA TensorRT-LLM-together with NVIDIA Hopper architecture computing resources to support data generation, model fine-tuning, inference optimization, and experimental evaluation. Leveraging NVIDIA AI ecosystem, Cathay FHC continues to strengthen its capabilities in financial-domain model development, data governance, and application validation.

In recent years, Cathay FHC has steadily expanded AI innovation across a wide range of financial scenarios, building scalable technological foundations spanning internal process optimization, customer service enhancement, financial knowledge understanding, and model governance. As financial institutions navigate an increasingly regulated environment, stringent data governance requirements, and rapidly evolving customer expectations, Cathay FHC remains committed to advancing AI research in a compliant, secure, and resilient manner.

Looking ahead, Cathay FHC will continue exploring long-context classification, advanced financial document understanding, and cross-scenario AI applications. By developing model training and deployment approaches tailored to the financial sector, the company aims to accelerate innovation and create more intelligent, efficient, and customer-centric financial services.