Teams that need to train models on proprietary or highly specialized tasks often start with data that already exists and then spend weeks labeling filtering and reshaping it to match the desired behavior. This process caps model quality at how well the available data aligns with the target task and is especially painful when the relevant signal lives in internal systems unstructured text or workflow logs that do not convert cleanly into a focused training set. Traditional synthetic‑data tools add another layer of work because they require a predefined schema task distribution and generation strategy before any rows can be produced.
Adaption Labs addresses this bottleneck with Invent a Dataset a feature that creates a ready‑to‑train dataset directly from a natural language description of the behavior you want the model to learn. No seed corpus no schema and no labeling guide are needed. The workflow is simple: call datasets.invent with one or more domain codes (such as medical or legal) optionally narrowed by subdomain codes specify the number of rows choose between instruction pairs or preference pairs and if desired enable language expansion. The call returns immediately with a status of running; you poll datasets.get until the state becomes succeeded or failed then download the rows as JSONL JSON CSV or Parquet files that you own and can use anywhere. An estimate mode lets you price the request without consuming credits and an idempotency key makes retries safe.
Generated dataset IDs feed straight into AutoScientist which co‑optimizes the data and the training recipe closing the loop from intent to trained model. Internal evaluations show AutoScientist delivers an average 35 % gain over configurations built by the lab’s own researchers across eight verticals.
By turning a plain description into portable training rows Invent a Dataset removes the upfront data‑gathering bottleneck lets teams iterate faster and focuses effort on model improvement rather than data preparation.
#AI #Product #MachineLearning #Data #LLM #AutoML