Only one in- and one out-of-distribution experiment is not convincing

[juripetersen]

The plan is:

• Three in-distribution workloads and two out-of-distribution workloads
• Train three sets of models on the respective data
• Run each in-distribution workloads
• Pick one model to run the out-of-distribution workloads