What if a training set could simply be written into existence? Adaption Labs is shipping a feature that does exactly that: type in the behavior a model should learn, and Invent a Dataset returns a structured dataset, with no source corpus, schema or labeling guide required.
The tool is available now in the Adaption app, and developers can call it through a Python SDK or a REST API. Output rows come back as portable files in JSONL, JSON, CSV or Parquet, ready for any downstream training pipeline. Generation itself runs on Adaption’s hosted platform and draws down prepaid credits.
The company’s critique of standard practice is blunt: teams usually start with whatever data already exists, then spend weeks labeling and filtering until it approximates the behavior they want. Quality ends up capped by how well that starting material matches the task. For proprietary or niche work, the real signal is often buried in internal systems, unstructured text or workflow logs that never convert cleanly into a training set.
Adaption also distinguishes its approach from earlier synthetic-data products. Those tools automate generation after a human has already fixed the schema, the task distribution and the generation strategy. Invent a Dataset begins one level earlier, at the behavior itself, and derives the schema and examples from there.
The underlying bet is that describing a job will soon be enough to manufacture the examples for it, turning dataset construction into a prompt-level skill rather than a data-engineering project.
