Overview TSBench is a large-scale synthetic benchmark dataset comprising 120,000 unique univariate time series, specifically designed to address the limitations of traditional statistical tests and the evaluation gaps in modern deep learning architectures, including Time Series Foundation Models and Large Language Models (LLMs). Dataset Structure & Organization The dataset is provided in two primary compressed formats, organized by complexity and labeling strategy: Generated_Data.zip: Contains time series characterized by a single primary property. These are ideal for fundamental property testing and baseline evaluations. Combinations.rar: Contains complex time series where multiple characteristics (e.g., trend + seasonality + volatility) coexist, providing a more challenging environment for advanced models. Data Labeling and Metadata: Folder-Based Labeling: The dataset follows a hierarchical directory structure where labels are assigned based on the sub-folder names. Granular Metadata: In addition to folder-based labels, each leaf directory contains a metadata file that stores precise generation parameters and specific series attributes for high-fidelity research. Technical Summary Total Size: 120,000 distinct time series. Sequence Lengths: Short (50–100), Medium (300–500), and Long (1,000–10,000). Key Properties: Stationarity/Non-stationarity, Deterministic/Stochastic Trends, Multi-frequency Seasonality, GARCH-based Volatility, Diverse Anomalies, and Structural Breaks. Primary Tasks: Stationarity Detection, Trend Classification, Anomaly/Break Localization, and LLM Benchmarking. Code and Resources The source code for data generation and the underlying framework is available at: https://github.com/ismailguzel/ts-stationary Funding & Acknowledgments This work was supported by: TÜBİTAK (The Scientific and Technological Research Council of Türkiye) under Grant No. 124F095. METU (Middle East Technical University) Scientific Research Projects Coordination Unit under Grant No. GAP-109-2023-11361.