During the past few decades, several studies have introduced various Subjective Video Quality Datasets (SVQDs). Most of these datasets are small in size with regard to the number of unique content and consist of relatively short video clips. Only a few recent SVQDs include videos longer than a minute in the total duration. Another drawback of current SVQDs is the fact that they only provide a single quality score for the quality of the whole video irrespective of its length. The given score also does not provide enough insight into different variations that occurred in the quality of the video and the type of utilized distortion and/or enhancement method. Even though some of the recent state-of-the-art SVQDs have collected real-time ratings through subjective tests along with the overall quality scores, their number and availability are limited for the research community. Furthermore, most of them employed similar approaches in collecting real-time ratings by using a rating bar. Since this method requires the observer to perform several types of tasks together, e.g. watching the video and moving the rating bar simultaneously, it may introduce distraction and uncertainty in the observers' ratings. Similarly, current Video Quality Metrics (VQMs) mainly provide a single (or multiple) global quality score(s) that correspond to the quality of the video in its entirety. Such a quality score does not provide any explainable detail about the video, nor information about the video quality at a specific time in the video. In this study, we introduce the new COSI Subjective Video Dataset (CSVD) which includes 30 videos in total created from 10 various indoor and outdoor scenes. The main distortion type being explored is the High Efficiency Video Coding (HEVC)/H.265 compression method. All of the videos are two minutes in length and include time-varying quality fluctuations. To this end, a framework was developed to simulate various gradual and fast increments and decrements in the quality of stimuli.