Loss curves: inspect evidence before diagnosis

Paired loss observations

Blue: training. Amber: validation. Solid: raw observations. Dashed: trailing mean. Scroll the table to inspect all columns.

Read comparable observations

This direct data tool inspects paired training and validation losses; it does not train a model. Enter JSON arrays with two to two hundred nonnegative finite numbers. Each pair must describe the same observation checkpoint and comparable loss definition, averaging convention and units. Index zero denotes the first supplied checkpoint. The horizontal axis uses equally spaced observation indices, not elapsed training time or dataset size. The default arrays are a transparent teaching example, not measured results. Training and validation share one vertical scale starting at zero, so their separation remains meaningful. Blue is training and amber is validation; the full table supplies the same information without relying on color.

Choose smoothing and inspect a point

Window is the number of observations in a trailing arithmetic mean. At the beginning, the window contains only available observations; there is no zero padding and no use of future data. Solid lines show raw data and dashed lines show means. A larger window can hide a spike and delays visible changes. It never changes the raw loss gap, raw minimum or stopping decision. Inspect index selects a zero-based row, highlights it in the table and moves the purple chart cursor. This is an immediate numerical inspector: results remain available indefinitely, with no artificial playback clock. Reset defaults restores both arrays, window one, patience two, minimum improvement zero and inspect index four.

Reproduce the default result

With training [4,3,2,1,0.5] and validation [4.2,3.2,2.2,2.4,2.6], the smallest validation loss is 2.2 at index two. At index four the validation-minus-training gap is 2.1. Set Window to three: the last training mean is 3.5 divided by three and the last validation mean is 2.4. The raw minimum remains 2.2. Change the final validation entry to 1.8: the minimum now occurs at index four, its raw gap is 1.3, and the default stopping policy no longer stops within the supplied observations. These outcomes follow directly from the arrays; nothing is randomly regenerated when a control changes.

Understand the stopping policy

Patience is an integer number of consecutive observations without sufficient improvement. Minimum improvement is a nonnegative absolute loss difference. The first observation initializes the accepted checkpoint. A later raw validation loss replaces it only when accepted loss minus new loss is strictly greater than Minimum improvement; equality does not qualify. An accepted improvement resets the counter. Otherwise the counter increases and the policy stops when it reaches Patience. The default stops at index four and retains checkpoint two. The policy freezes after stopping, although later supplied observations remain in the chart and global minimum calculation. This makes the difference between a hypothetical stopped run and hindsight explicit. No framework compatibility is implied.

Test ties and avoid unsupported diagnoses

Use validation [3,2,2,1] with patience one and minimum improvement zero. The tie at index two stops the policy with checkpoint one, while the lowest supplied loss is one at index three. A rising gap can be consistent with overfitting but does not uniquely identify its cause. Check held-out data, metric normalization, distribution shifts, train versus evaluation behavior and repeated runs before changing a training system. A spike alone does not prove a bad batch or a learning-rate bug. This tool cannot estimate uncertainty, test-set performance or an optimal hyperparameter. Repeatedly choosing settings on the same validation set can itself overfit the selection process.

Check boundaries and exact values

Negative losses, nonnumbers, mismatched lengths, invalid JSON, empty numeric controls and out-of-range indices are rejected. Window cannot exceed the array length. Patience may exceed the available observations, in which case no stop need occur. Loss and minimum improvement are bounded at one million; this is a teaching input limit. Full precision calculation exposes every row and stopping counter in JavaScript double precision; table decimals are rounded. Scroll the table horizontally on a phone to inspect both means. Shared notes can record assumptions and comparisons, while Capture baseline retains a visible experiment snapshot. Neither action uploads a training log.