Here's the code that generates the belgium and usa datasets:
https://github.com/ge0ffrey/vrp-dataset-generator

For both belgium and usa, we want to have a timewindowed with real road-time data.
(generating real road-time data for USA is a pain due to the sheer size of the geo file.)

Requirements:

• A feasible solution shouldn't be too hard (a good CH should suffice to get it feasible)
• Normalization on seconds: both readyTime/dueTime and distances should be in seconds
• A nice way to represent on the time windows clock rendering somehow. For belgium datasets this can be done by using a total window of 12 hours. For the USA datasets it's impossible to have a feasible solution in 12 hours as it takes much longer to travel from atlanta to los angelos and back.
• different scales
• realistic number of stops per vehicle. 100 stops might not be feasible, unless the total time window does span days?

The current datasets generated by my code above don't fulfill all these requirements. The belgium-road-time-tw's are even unfeasible.