Add hyperparameter sweep

[nilsvselte]

• On one type of training

• Smaller vs normal vs larger model

• Attention heads etc.

[nilsvselte]

• Just on linear class (assume works on both)
• One type of training