Training examples, ready for review.

By Axeron AI

Relay Lab connects the preparation of synthetic training examples with a visible handoff to ModelForge. This recorded workflow follows a manufacturing data lead from a plain-language request to sixteen checked examples selected for a training experiment.

For a data lead, a useful starting point is a task whose inputs, expected answer and review method are clear. That makes it possible to inspect what has been generated and decide whether the examples are suitable for the next step.

The wider industry context makes this preparation relevant. Deloitte’s 2025 survey of 600 executives from large manufacturing companies identifies data readiness and workforce adoption as practical concerns in implementing smart manufacturing. Our approach here starts with a small, checkable task that a reviewer can follow from beginning to end. Deloitte’s 2025 Smart Manufacturing and Operations Survey.

One task, defined precisely.

The demonstration asks for calculation examples about accepted production pieces. Every question must include the production rate, working hours and percentage rejected. Every answer must return the whole number of pieces accepted.

In Synthetic Data, the data lead names the set, describes the task, chooses Distill, selects the question and answer models, and sets the example count and spending cap. Generate starts the work. The application displays the arriving pairs, progress and reported spend.

This makes the request concrete before generation begins. The reviewer knows the information each question should contain and the result each answer should produce.

Every answer, open to inspection.

The saved set contains sixteen examples. Opening a pair reveals the question, reference answer and supporting context.

One question describes a machine producing 200 pieces an hour for eight hours, with 18 percent rejected. Total production is 1,600 pieces. The accepted share is 82 percent, giving 1,312 accepted pieces. The displayed answer matches that calculation. A second example uses 350 pieces an hour, eight hours and 12 percent rejected, giving 2,464 accepted pieces.

All sixteen answers in this demonstration were checked against the quantities in their questions before the set moved forward. Distill retains generated answers; the arithmetic check supplies the separate evidence that these particular answers are correct.

The practical benefit is inspectability. The data lead can see the examples being prepared and apply a clear acceptance rule to their contents.

The same dataset, in the training form.

Selecting Use in ModelForge carries the set into the next step. ModelForge displays the same dataset name, sixteen rows and three fields. The reviewer can confirm that the selected training input is the set just inspected.

Synthetic Data and ModelForge therefore connect two concrete activities in this workflow: preparing examples and selecting them for a training experiment. The demonstrated result is the reviewed set and its visible handoff. Any resulting model’s quality is evaluated separately on new examples after training.

A practical pilot, with agreed checks.

A department pilot can begin with one supported task, an agreed answer format and a reviewer who knows what a useful result looks like. Keep separate examples for later evaluation, record the dataset and model choices, and measure the effort needed to produce an acceptable result, including review and correction.

For the sponsor, that creates a clear discussion: which task is worth preparing data for, what should its examples cover, and what evidence will justify the next investment?

Explore the recorded workflow, then work with Axeron AI to define a pilot around a task your team can verify.