T-CVRP v1.0 is a synthetic benchmark dataset of 10,000 travel-time-based Capacitated Vehicle Routing Problem (CVRP) instances produced within the SmartDelivery Marie Sklodowska-Curie Postdoctoral Fellowship (Grant Agreement No. 101110022). The dataset was created to address a methodological gap in the CVRP literature. Standard benchmark collections typically use Euclidean distances as routing costs, whereas realistic last-mile delivery optimization depends primarily on travel time. T-CVRP v1.0 therefore provides benchmark instances in which routing costs are represented by explicit travel-time matrices rather than by Euclidean distances. The dataset was generated with a dedicated Python pipeline that starts from CVRP-style benchmark structures and applies a parameterized travel-time model. The model incorporates multiple effects relevant to urban logistics, including heterogeneous road conditions, stochastic congestion, weather scenarios, and additive stop-and-go delays such as traffic lights and intersections. The generated matrices are post-processed to preserve symmetry and to enforce triangle inequality for solver compatibility. Public Distribution Two public distributions of the dataset are available: - A structured archive, intended for detailed inspection and reuse. In this version, each instance is accompanied by a descriptive statistics file.- a flat archive, intended for users who only need the benchmark instance files and do not require the directory structure or per-instance statistics. All benchmark instances are provided in VRPLIB-compatible `.vrp` format with explicit full travel-time matrices. In the structured archive, each instance is accompanied by a `.csv` statistics file reporting descriptive information useful for reproducibility and later feature engineering. The benchmark covers customer sizes from 30 to 200, with step 5. The generation pipeline was designed to produce a balanced dataset across multiple structural and travel-time-related factors, including: - depot positioning- customer spatial distribution- demand profile- average route-size class- weather condition The dataset is fully synthetic. It does not contain personal data, real operational fleet traces, customer information, or other identifiable or confidential data. Example structure of the structured archive The structured archive is organized to make navigation easier. Depending on the exact release package, the internal folders group the instances by generation parameters such as size class and scenario settings. Each instance is paired with a statistics file. A typical layout is illustrated below: T-CVRP_v1.0_dataset_directories/├──  │ ├──  │ │ ├── .vrp │ │ └── _stats.csv Example instance name: LDG30_1265_rain_30_0001 In this naming pattern: - '30' indicates the number of customers- 'rain' indicates the weather scenario- the remaining tokens distinguish the generated configuration and unique instance identifier Statistics files The statistics file associated with each instance summarizes the main properties of the generated problem instance. Depending on the generation settings, the file may include: - instance name- weather condition- layout setting- generator parameter values- travel-time and Euclidean-distance summaries- demand summaries- travel-time-to-distance indicators- congestion-related indicators These files were produced to support reproducibility and future machine-learning-oriented feature extraction. Related project materials The GitHub repository associated with the project contains the scripts and documentation used to generate and interpret the dataset, including: - the travel-time-based CVRP instance generator- the bash-script generator used to automate large-scale generation- a travel-time-based CVRP solver- a prototype simulator for the SmartDelivery "sixth sense" route-to-driver assignment module- technical documentation and and README files Link: https://github.com/gaetano78/SmartDelivery The original SmartDelivery plan also envisaged the creation of a derived machine-learning dataset containing discriminative features and best-solver labels for algorithm selection. This follow-up dataset was not completed before the early termination of the MSCA action. However, the released benchmark instances and per-instance statistics were explicitly designed to support this future extension. License The public dataset is distributed under the CC BY 4.0 license. Citation If you use the dataset, please cite the Zenodo record: - La Delfa, Gaetano Carmelo. T-CVRP v1.0 - 10,000 balanced, travel-time-based CVRP instances. Zenodo.