qsWDFGHN/FIN-VQA

FinVQA-Chart是一个突破性的大规模基准数据集，旨在评估视觉语言模型（VLMs）在复杂金融推理任务上的能力。与现有主要关注通用图表理解或金融文本分析的数据集不同，FinVQA-Chart独特地结合了跨越5年（2020-2024年）的关键市场时期（COVID崩溃、通货膨胀激增、美联储收紧）的真实世界金融数据，需要同时理解视觉模式、数值数据和市场背景的多模态推理，具有35,000多个专业检测到的蜡烛图模式和信号的技术分析深度，以及跨越7个不同推理类型和3个难度级别的141,266个问题的规模和多样性。这个数据集通过要求模型整合视觉图表解释、数值推理、时间理解和特定领域的金融知识来推动VLM能力的边界，即使是最先进的模型，这些任务仍然具有挑战性。

FinVQA-Chart is a groundbreaking large-scale benchmark dataset designed to evaluate Vision-Language Models (VLMs) on complex financial reasoning tasks. Unlike existing datasets that focus on generic chart understanding or financial text analysis, FinVQA-Chart uniquely combines real-world financial data spanning 5 years (2020-2024) across critical market periods, multi-modal reasoning requiring simultaneous understanding of visual patterns, numerical data, and market context, technical analysis depth with 35,000+ professionally detected candlestick patterns and technical signals, and scale and diversity with 141,266 questions across 7 distinct reasoning types and 3 difficulty levels. This dataset pushes the boundaries of VLM capabilities by requiring models to integrate visual chart interpretation, numerical reasoning, temporal understanding, and domain-specific financial knowledge—tasks that remain challenging for even state-of-the-art models.