Adaption Labs Launches Invent a Dataset to Generate Training Data from Task Descriptions

Adaption Labs argues that typical dataset workflows begin with existing data and require weeks of labeling and filtering, which limits model quality to how closely available data matches the intended behavior. For proprietary and specialized tasks, relevant signals usually reside in internal systems, unstructured text, or workflow logs and rarely convert cleanly into a focused training set. The research team also distinguishes Invent a Dataset from existing synthetic-data tools, which automate generation only after a human has already defined the schema, task distribution, and generation strategy. Invent a Dataset is said to start one level earlier, at the behavior itself.

The API mechanics are documented: a single call to datasets.invent creates the dataset and starts generation, returning status running; teams poll datasets.get until the status is succeeded or failed. Domain codes serve as the primary control; developers fetch current codes with datasets.invent_domains instead of hardcoding them, then pass values such as medical, which can be narrowed by qualified subdomain codes like medical.symptoms_diagnosis. At least one domain or subdomain is required, and multiple domains can contribute to the same run.

Two output formats are supported. The default instruction_dataset produces prompt-completion pairs for supervised fine-tuning; preference_pairs outputs chosen and rejected completions for preference-based training such as DPO. Three parameters are highlighted for production use: estimate=True prices the request without charging, prompt accepts up to 10,000 characters to steer row content, and idempotency_key prevents duplicate runs on retries. Row counts are subject to a per-launch plan limit.

Language expansion runs in two modes: translate produces a new row variant per target language, while localize produces variants per country-language pair with locale-specific wording. A sample_rate between 0.01 and 1 controls the fraction of invented rows expanded, and credits are billed on expanded output rows, not the original count. Unsupported codes return a 400 error with sample valid values.

Invent a Dataset is positioned as the first half of a loop: its dataset ID passes directly to autoscientist.create, which co-optimizes the data and training recipe against the objective. AutoScientist, launched in May 2026, is the training-side counterpart to the Adaptive Data pillar. According to MarkTechPost, Adaption reports that AutoScientist beats training configured by its own research staff by an average of 35%, moving win rates from 48% to 64%; these figures come from in-house domain-specialized evaluations across eight verticals, with dataset sizes ranging from 5,000 to 100,000 rows on architectures offered by Together AI.