Improve with evidence
Generate for sparse data
Fill specific coverage gaps with reviewed synthetic examples and protected evaluation data.
Synthetic examples help explore variations of a reviewed situation when real examples are sparse. Start from trusted real evidence, keep the governing requirements fixed, and review every target you accept.
Identify a specific coverage gap#
Suppose you have reviewed an authorization-test deletion, but have few nearby examples. Useful coverage goals include a legitimate implementation repair, a subtler attempt to weaken an assertion, and a record too incomplete to support a decision. These distinctions test both useful capability and avoidance of divergence.
Choose a reviewed real dataset compatible with the component you want to adapt. Generation uses examples from its training-side source families. Held-out cases remain reserved for evaluation, and synthetic examples cannot become the anchors for another generation cycle.
If your missing coverage is an unfamiliar customer task or environment, collect real examples as well. Changing names or paths in one session does not establish representative coverage of that workload.
Generate, grade, and review#
- Prepare drafts. Where Data offers Generated coverage, select the source dataset and request a bounded number of examples. The preparation request defaults to 4 and accepts 1 to 20.
- Keep the contract fixed. The generator varies new evidence under the reviewed task and constitution rather than rewriting the requirements to make a target acceptable.
- Inspect independent grading. A separately configured grading model checks whether each proposed target is supported by its evidence. Unsupported drafts are not eligible for acceptance.
- Review the exact target. Approve the displayed target only when you agree with the evidence and label. A changed target needs a new independent review.
- Save the reviewed selection. Accepted records retain synthetic provenance and their connection to the source examples. They enter a separate training-only dataset.
Inspect the desired behavior, proposed action, and evidence together. Reject a draft that invents a protected test, claims an unobserved result, or loses useful task context. Model agreement is a review input, not a substitute for your judgment.
Keep generated and real evidence distinct#
| Data | Use in a platform training batch |
|---|---|
| Reviewed real training examples | Establish observed task coverage and provide generation anchors. |
| Reviewed synthetic examples | Supplement training coverage within the configured mix cap. |
| Independent real held-out examples | Evaluate behavior without training on the same source families. |
Platform training batches cap the synthetic share at 50%, or a lower configured limit. Synthetic examples do not satisfy the minimum real-record or real-family requirements, and cannot replace real held-out evidence. Review readiness and calibration before preparing a batch.
Preparation limits and availability#
Coverage preparation depends on the workspace’s configured generator, independent grader, background worker, and compatible dataset. The coverage service permits up to 2 pending preparations and 10 new preparations per hour per workspace. If a draft is held, inspect its recorded reason before making another request.
Generating and reviewing data is separate from activating a trained version. The selected connection determines whether it can use customer artifacts; see versions and adaptation. Retention and source access continue to apply throughout preparation and review.