Each pattern tells you something different about the training run.

What you're seeing

Plot training (blue) and validation (green) loss over steps.

★ KEY TAKEAWAY
Loss curve shape tells you everything: healthy = smooth descent, spike = bad batch, divergence = bug, train≪val = overfitting.
▶ WHAT TO TRY
  • Switch between the four modes.
  • In production, watch for these patterns and intervene early (lower LR, restart, stop, more regularization).
← Back to Transformer Math & CPU SLM