Run Status
Per-run health at a glance. Runs whose telemetry ends mid-epoch are marked IN PROGRESS; resume seams count process restarts detected via elapsed-time resets.
Summary
Loss Curves
Running Avg is cumulative from step 1 — it lags Batch Loss by including early high-loss steps. Window Loss tracks the recent moving average. Dashed vertical lines mark resume seams (run-colored) and epoch boundaries (grey, Epoch Bands mode).
Training Loss Over Time
Smoothed Loss (EMA overlay)
Eval Loss
Best Eval Progression
Train–Eval Loss Gap
Bits per Byte (BPB)
Loss Derivatives
Central differences of Window Loss per contiguous epoch segment, scaled per 100 steps. Velocity near zero signals a plateau; acceleration spikes flag turbulence or resume seams.
Loss Velocity ΔL/Δstep
Loss Acceleration Δ²L/Δstep²
Accuracy
Eval metrics come from held-out eval rows only. ★ star markers flag evals that set a new all-time best loss.
Top-1 / Top-5 Accuracy (%)
Per-Category Eval Accuracy (%)
Gradient & Learning Rate
Gradient Norm (Pre/Post Clip + RMS)
Gradient Norm — Log Scale (spike view)
Learning Rate Schedule (WSD)
Teacher Forcing Schedule
Clip Scale Over Training
Training Throughput
Sequences per Second
Tokens per Second
Total Tokens Trained
Batch Time (ms)
Throughput vs. Batch Time
Forward / Backward / Clip / Optimizer Timing (ms)
Memory & Weight Norms
Weight norms interpolate between real checkpoint measurements where GPU telemetry gaps exist.
Memory RSS Over Training
GRU Layer Weight Norms
Aux Weight Norms (RMSNorm / Classifier / Vocab)
Best Checkpoints
Top checkpoints across all runs by eval Top-1, then lowest eval loss. ★ marks evals that set a new record at the time.
Eval History
Last 8 held-out evaluations per run. Δ is the change vs the same run's previous eval.