Platform / nQube · Training Center
Fifty-four checks and a price, before the first GPU hour. A person says go.
Product screens recreated with synthetic data.
Training a model is the most expensive thing a workflow can do. Five to ten hours of GPU time is ordinary, and most of the failures show up at the end: a feature column that shifted, a label that leaked the future, a window too short for the pattern, a learning rate that diverges on step three. The Training Center puts those failures in front of the money.
The run is six container stages on one canvas. Prepare, features and sequences run on CPU and each writes an artifact the next stage downloads, stamped with the configuration hash. Then the pre-flight runs where the training will run: it re-derives the feature columns, builds synthetic sequences, takes five real optimizer steps, checks that batch and single-row inference agree, and prices the run from the step time it just measured. GO goes to a person with the estimate. NO-GO ends the run at no further cost, with the failed checks and a suggested fix.
The model is never the gate. The pre-flight is arithmetic, and the approval is a human review item like any other on the platform. The GPU stage starts only behind both, re-uploads its best checkpoint after every improving epoch, and the evaluation reports error in the target’s own units next to a naive baseline, so the number means something.
The Training Center is not a notebook and it is not a model zoo. It runs the training graph you drew, with your data, under the gates you placed. The bundled dataset is synthetic — a shape that exercises every stage, not a measurement of anything.
A data team that trains models and pays for the GPU, and the person who signs off on the spend. You do not need to read a training log to know whether a run was allowed to start, what it was expected to cost, and what it returned.
Platform. The canvas, the pre-flight block and the Training Center tab are in nQube. GPU stages run on the GPU capacity provider, metered per run, on plans with compute turned on.
Pricing →Synthetic example. Not a customer.
You open the Training Center and press Run. The three data stages finish in under a minute. The pre-flight panel reveals its checks one at a time — fifty-three pass, one advisory. The cost strip reads an estimate of a few dollars for twenty minutes against a ten-dollar budget. A Human Review item appears with that estimate in its summary line; you approve it. The GPU stage starts, the loss curve draws itself epoch by epoch, and the evaluation lands: error in kilowatts next to the naive baseline, R² on the held-out windows, the six features the model leaned on. The report is in Knowledge. On a different afternoon, the same panel says NO-GO on the first check that matters, and the GPU is never touched.
The checks that used to surface after five hours surface in ninety seconds, on the data the run will actually use.
The estimate comes from a measured step time, not a guess, and a person approves it as a review item.
The best checkpoint is re-uploaded after every improving epoch. A timeout or a cancel still leaves a model behind.
Evaluation reports in the target’s own units next to a baseline. Improvement over the naive answer is the headline, not the loss.
The CUDA device, the disk, the config hash. Every upstream artifact is present, readable, and built from the same configuration — a stale artifact cannot be trained on.
The feature list is re-derived from the config and must equal the artifact’s. The matrix width must equal that count. The target column comes first. The calendar features are there.
Features and sequences are rebuilt on a synthetic frame: no NaN or infinity, every feature varies, the label is exactly horizon steps after the window, nothing leaks across the train/validation boundary.
The model builds, the output has the right shape, the selection weights sum to one, the loss is finite and does not diverge, gradients reach every layer.
Batch equals single-row inference. The head reads the last valid timestep. The window is longer than the longest rolling span; there are enough windows per feature.
Measured milliseconds per step, times the steps the config implies, at the GPU hourly rate: an estimate in dollars and minutes, checked against the cap you set.
Six container stages, one image, a Pre-flight check block in front of the GPU stage, a Human review behind it. The seeded example is one click to duplicate.
The bundled synthetic sample, or your own CSV or Parquet by URL with the timestamp, target and feature columns named in the config.
The stage rail, the checklist, the cost strip, the loss curve — on the Training Center tab and on the canvas.
GO waits for you in Human Review with the estimate. NO-GO explains itself. Cancel stops the GPU and keeps the checkpoint.
Platform Start is the self-serve canvas. A first pipeline we stand up with you is delivery, priced apart from the subscription.