Adaption Labs lets developers generate training datasets from a task description alone

Adaption Labs this week released Invent a Dataset, a feature that builds a training-ready dataset directly from a written description of the behavior a model should learn. Users do not need to supply a seed corpus, a schema, or a labeling guide beforehand.

The company positions the tool against how most machine learning teams currently work. Typical projects start with data that already exists, then spend weeks labeling, filtering, and reshaping it to approximate a target task, a process Adaption says caps model quality at how closely that raw data matches the intended behavior. For specialized or proprietary tasks, the useful signal often sits in internal systems or unstructured logs that don’t convert cleanly into a training set. Adaption also draws a line against existing synthetic-data tools, which it says still require a person to define the schema and generation strategy before automation begins; Invent a Dataset instead starts one level earlier, at the behavior description itself.

Mechanically, a single call to datasets.invent creates a dataset and starts generation, returning immediately with a status of “running.” Developers then poll datasets.get until the job reports “succeeded” or “failed” before downloading rows as JSONL, JSON, CSV, or Parquet files. Domain codes, fetched through a separate datasets.invent_domains call rather than hardcoded, control subject matter: a broad code like “medical” or a narrower one such as “medical.symptoms_diagnosis,” with at least one domain or subdomain required per run. Output comes in two formats: instruction_dataset, the default, produces prompt-completion pairs for supervised fine-tuning, while preference_pairs produces chosen-and-rejected completions for preference-based methods such as DPO.

Adaption built in a few production safeguards. Setting estimate=True prices a request in credits without charging or creating anything. An idempotency_key of up to 255 characters makes network retries safe by returning the original dataset rather than launching a duplicate run. A prompt field of up to 10,000 characters can steer what the generated rows are about, and an optional language_expansion setting can translate or localize a sample of rows into other languages, with credits billed on the expanded row count rather than the original.

Feeding into AutoScientist

The dataset ID from Invent a Dataset passes directly into autoscientist.create, the training-side tool Adaption launched in May 2026 that optimizes the data and the training recipe together against a stated objective. Adaption reports that AutoScientist beat training setups configured by its own research staff by an average of 35%, raising win rates from 48% to 64%. Those figures come from in-house, domain-specialized evaluations across eight verticals, using dataset sizes between 5,000 and 100,000 rows on architectures available for fine-tuning through Together AI. Adaption has not published independent or third-party benchmarks of these results.

Generation itself runs only on Adaption’s hosted platform and draws on account credits; no self-hosted option is documented. Whether third-party evaluations confirm Adaption’s in-house benchmark numbers, and how the tool performs on tasks outside the eight verticals already tested, will become clearer as more teams adopt it.