loss-curve
Lint your training logs after the run: NaN, divergence, loss spikes, overfitting onset and step bugs in trainer_state.json / CSV / JSONL / TensorBoard event files, with CI-friendly exit codes.
