lr starts at 0.5, already 100% of the run's peak (0.5) -- no ramp-up phase visible in the log.

initial_lr=0.5   peak_lr=0.5

• Add LR warmup (e.g. linear warmup over the first 1-10% of steps) -- it's a common fix for early spikes/instability.
• If warmup is configured but not logged until after it finishes, this may be a logging artifact -- check the logging interval.

train/loss jumped to 0.5146 at step 203, 7.4x the local median-absolute-deviation scale.

peak_step=203   peak_value=0.5146   robust_z=7.419   window=21   threshold=6

• Lower the learning rate or add/extend LR warmup.
• Clip gradients (or lower the existing max-norm).
• Check the batch at this step for a data or tokenization bug (e.g. an unmasked pad token, a corrupted shard).
• If it self-recovers within a few steps and doesn't recur, it may be benign -- confirm against the divergence detector's verdict.

Best eval/loss was 0.146 at step 230; it's since risen to 0.1941 (32.9%, slope p=0.0245).

best_step=230   best_value=0.146   current_value=0.1941   relative_rise=0.3293   slope_p_value=0.02446

• Use the checkpoint at step 230 (best eval/loss), not the last one.
• Add or increase regularization: weight decay, dropout, data augmentation.
• Reduce model capacity or train for fewer steps / add early stopping.
• Get more training data, or de-duplicate against the eval set if overlap is possible.