Constructing Synthetic Instruction Datasets for Improving Reasoning in Domain-Specific LLMs: A Case Study in the Japanese Financial Domain

Researchers developed a method for creating synthetic instruction datasets to improve domain-specific LLMs, demonstrating with a 9.5 billion token Japanese financial dataset. The approach enhances both domain expertise and reasoning capabilities, with models and datasets being open-sourced for broader use.

• →A new method constructs high-quality synthetic instruction data for domain-specific LLM training starting from domain vocabulary.
• →Researchers created a 9.5 billion token financial instruction dataset with Chain-of-Thought reasoning traces.
• →Evaluation showed performance improvements over baseline models on financial benchmarks.
• →The study examined the impact of reasoning trace length on model performance and identified limitations.
• →Models and datasets are being open-sourced on Hugging Face for community access.