Establishing a high-quality functional Near-Infrared Spectroscopy (fNIRS) emotional dataset under emotion-inducing conditions is of significant research importance. In this paper, we present a multimodal affective dataset based on fNIRS, comprising recordings from 28 participants observing five categories of emotional videos (pride, happy, neutral, fear, sad). The dataset includes fNIRS data, Galvanic Skin Response (GSR) data, Photoplethysmographic (PPG) data, and facial expressions data. This study explores the variations and distinct activity patterns of oxyhemoglobin activation under five emotional states. To facilitate classification, deep learning and machine learning models are utilized to perform two-class and five-class classification experiments, both for individual subjects and across multiple subjects, yielding results for both subject-specific and cross-participant models. The dataset and source code are publicly available to promote research on emotion recognition using fNIRS data.