Vietnamese Financial Reports and Evaluation Dataset for Agentic Financial Analysis

This repository contains a comprehensive benchmark for agentic financial analysis over Vietnamese corporate financial reports.The dataset includes publicly available annual financial reports in PDF format collected from Vietnamese enterprises, together with multiple evaluation datasets designed for benchmarking retrieval, reasoning, and financial analysis capabilities of AI systems.The benchmark consists of:100 high-quality financial question–answer pairs generated using Gemini 2.5 Pro and manually reviewed by human annotators to establish reliable ground-truth annotations.An additional collection of more than 300 financial QA pairs, provided as supplementary evaluation data for both LLM-based and SLM-based systems.An open-ended financial reasoning benchmark containing 119 question–answer pairs constructed from the 2025 annual reports of 8 Vietnamese companies representing different business scales.

Each sample requires the system to analyze financial statements, integrate contextual information, and produce comprehensive analytical responses rather than simple fact retrieval.The repository also provides the complete financial reports used in the experiments and the evaluation results of different methods under the RAGAS framework, enabling reproducible comparison across retrieval-augmented and agentic financial analysis systems.