Your layout model has five pre-training objectives comb...

Weights are meaningless unless the losses share a common range, so normalization is the first step. For validation, use a per-corpus hypothesis to determine which objectives the data can actually supervise. Learned weights are often unreliable as they optimize for the training objective rather than your intent, potentially favoring shortcuts over structural learning. Ablation studies are the only reliable way to confirm that the model is learning the intended structure.