Practical Exploration of Small Language Models in Vietnamese Financial Numerical Reasoning

This article focuses on the research of Small Language Models (SLMs) in Vietnamese financial numerical reasoning tasks. It constructs the ViNumQA dataset as an evaluation benchmark, and verifies the practical value of small models in low-resource scenarios through prompt engineering (e.g., Chain-of-Thought), self-assessment mechanisms, and domain optimization strategies. The study shows that the optimized 7B-parameter model can achieve satisfactory numerical reasoning capabilities, providing a reference for AI applications in resource-constrained environments.

Vietnam, as a fast-growing economy in Southeast Asia, is accelerating its financial digital transformation. However, Vietnamese financial text processing faces multiple challenges: significant differences in language characteristics from English, unique numerical expressions, strong domain professionalism, and scarcity of annotated datasets. This study aims to explore the feasibility of small language models in this scenario.

The research team constructed the ViNumQA dataset, which features diverse sources (financial reports, news, etc.), rich question types (direct extraction, comparative calculation, etc.), difficulty stratification, and professional annotation. The included reasoning types are direct extraction, arithmetic operations, comparative analysis, trend inference, and multi-step reasoning.

Advantages of choosing small models (1B-7B parameters): low deployment cost, fast inference speed, flexible customization, and privacy protection. The models evaluated in experiments include PhoGPT (Vietnamese-optimized), SeaLLM (Southeast Asian multilingual), Qwen-Chat, and Llama-2-Chat.

1. Prompt Engineering Optimization: Zero-shot (testing basic capabilities), few-shot (3-5 examples are optimal), Chain-of-Thought (guiding step-by-step reasoning to improve accuracy); 2. Self-Assessment Mechanism: Verifying answer rationality, cross-validation, confidence evaluation, and triggering manual review for low-confidence results; 3. Domain-Specific Optimization: Building a financial terminology dictionary (term mapping, abbreviation expansion, etc.), numerical normalization (handling Vietnamese-specific numerical expressions).

Key findings: Chain-of-Thought improves accuracy by 15-20%; 7B models with optimization approach the performance of 13B models; self-assessment further enhances accuracy; performance is significantly improved after domain fine-tuning. Error classification: Numerical positioning (35%), calculation errors (28%), unit confusion (18%), etc. Cross-model comparison:

Model	Zero-shot	Few-shot	CoT	+Self-assessment
PhoGPT-4B	42%	51%	58%	63%
SeaLLM-7B	45%	54%	61%	66%
Qwen-7B	48%	56%	64%	68%
Llama-2-7B	38%	47%	52%	57%

Practical insights: Data first (high-quality datasets are the foundation), leverage of prompt engineering, more practical small model optimization, self-assessment enhances reliability. Financial applications: Automated financial report analysis, investment research assistance, compliance checks, customer service. Limitations: Small dataset size, limited model generalization, challenges in complex reasoning. Future directions: Expand datasets, multimodal fusion, real-time learning, cross-language transfer.