Kasikorn Business-Technology Group (KBTG) is pushing the boundaries of AI with THaLLE, a powerful Thai-language Large Language Model (LLM) designed specifically for the financial sector. This innovative AI model promises to revolutionize financial services and empower Thais with deeper financial knowledge.
THaLLE, which stands for Text Hyperlocally Augmented Large Language Extension, was born out of necessity, said Tawunrat Chalothorn, Senior Research Engineer at KBTG.
“We explored existing LLMs, but none fully met our needs. So, we built THaLLE as an in-house solution for Kasikornbank,” said Tawunrat.
The team focused on analyzing common customer questions and internal bank data to train THaLLE to provide accurate and efficient answers related to finance and investment. This approach ensures that THaLLE is tailored to the specific needs of Thai banking customers.
A Collaborative Development Approach
The development of THaLLE was a collaborative effort by NLP experts at KBTG Labs, who formed three specialized teams:
Data Team: Responsible for preparing data for AI training, including converting it into suitable formats.
Training Team: Focused on training the AI model using techniques like pre-training and fine-tuning to optimize performance.
Evaluation Team: Rigorously assessed the AI’s accuracy and ability to handle real-world scenarios, making the final decision on deployment.
“We’ve also developed key technologies to enhance THaLLE’s capabilities. These include a data retrieval system using Learning to Rank techniques and Ontology Population for graph-based answer extraction. These advancements have been recognized by the global NLP community,” said Tawunrat.
Navigating the Fast-Paced World of AI Development
Building an AI model in today’s rapidly evolving technological landscape is no easy feat. The THaLLE team at KBTG faced the challenge of keeping pace with constant advancements in AI, machine learning, and LLMs, where new research emerges almost every week, said Danupat Khamnuansin Advanced Research Engineer at KBTG.
“We had to be agile and adaptable,” said Danupat, recalling a situation where a more advanced model appeared just as the team was preparing to launch their own. “This required us to adjust our strategy and resources to stay competitive,” said him.
To navigate this dynamic environment, the team prioritized practicality and flexibility in THaLLE’s development. The model needed to be functional, adaptable, and open to incorporating new advancements in AI and machine learning. They also focused on ensuring compatibility with other models, regardless of the underlying technology.
Staying at the forefront of AI research was crucial. With generative AI being a relatively new field, KBTG assembled a dedicated team to track emerging technologies and research, ensuring THaLLE remained cutting-edge.
Balancing Performance and Resource Consumption
Another challenge was managing the computational demands of LLMs. These models require significant server resources, which can be a constraint in Thailand’s business and industrial landscape. KBTG addressed this by focusing on smaller and medium-sized models tailored for specific applications.
Currently, THaLLE is built on the Qwen 2.5 model, with 7-8 billion parameters. This strikes an optimal balance between performance and resource consumption, making it suitable for various financial and banking use cases.
THaLLE: A Powerful and Versatile Thai-Language AI
THaLLE has evolved significantly since its initial release. Originally an English-only model, it now functions in both Thai and English, addressing the growing need for a powerful financial AI that understands the nuances of the Thai language. KBTG has rigorously tested and refined THaLLE in various banking scenarios, ensuring its effectiveness in real-world applications.
Tawunrat said that ethere ar two key strengths that set THaLLE apart included exceptional performance and open-source collaboration.
THaLLE has demonstrated its advanced understanding of financial concepts by passing the internationally recognized Chartered Financial Analyst (CFA) exams. This achievement distinguishes it from other financial AIs.
THaLLE’s open-source nature encourages collaboration and allows developers to enhance its capabilities for specific purposes. This opens up possibilities for creating specialized tools like investment assistants, stock market Q&A systems, and financial chatbots.
Expanding the Horizons of Thai-Language AI
KBTG envisions a future where THaLLE’s capabilities extend beyond finance and banking. The company aims to foster a broader ecosystem (THaLLE Ecosystem) that impacts education and other industries seeking to leverage AI.
One example is KBTG’s collaboration with AI Singapore and Google Research on Project SEALD, which focuses on advancing LLM development in Southeast Asia.
KBTG is committed to driving the development of advanced Thai-language LLMs, led by Thai talent. The goal is to create models that are equal to or better than existing LLMs, empowering Thais and improving society through broader applications of AI technology.
