You're training a small lead-scoring model and the loss log reads: 36 -> 220 -> 4100 -> 91000 -> inf. Each step is bigger than the last. What's almost certainly wrong, and what's the fix?