Fix fine-tuning eval with prediction probabilities

[th0rntwig]

The addition of prediction scores causes ValueError: too many values to unpack (expected 3) during evaluation when fine-tuning. This PR solves this by handling predictions separately.

[urchade]

Thanks!