You have graded labels 0–4. Defend the choice between s...

Rule out the regression head first and for the right reason: the grades are ordinal, not interval, so squared error assumes a metric no annotator supplied. Then state what a flat softmax costs - exchangeable classes, no notion that 3 is nearer 4 than 0 is - and what ordinal thresholds buy: one direction, K − 1 monotone cuts, a continuous score with no ties, and gradient sharing between adjacent grades. Recommend multi-class with expected-relevance scoring when the model is gradient-boosted trees, an ordinal head when the scale is long or noisy.