Adaption Labs shipped something this week that changes how synthetic training data gets made. Invent a Dataset takes a plain-language description of the behavior you want a model to learn and generates a structured, training-ready dataset. No seed corpus. No predefined schema. No labeling guide. You describe the target behavior, and the system produces rows you can download as JSONL, JSON, CSV, or Parquet.
Mechanics are straightforward. A single datasets.invent call kicks off asynchronous generation, which you monitor by polling datasets.get until status flips to succeeded or failed. Domain codes are the main steering mechanism — you fetch current values via datasets.invent_domains rather than hardcoding them, and you can pass broad domains like medical or drill down with qualified subdomains like medical.symptoms_diagnosis. Multiple domains can contribute to one run. Output comes in two flavors: instruction_dataset for supervised fine-tuning with prompt-completion pairs, or preference_pairs for DPO-style training with chosen and rejected completions.
Three production-minded parameters stand out. estimate=True gives you an exact price quote against available credits before you commit anything. prompt accepts up to 10,000 characters to steer what the rows actually cover. idempotency_key makes retries safe — a failed network call won’t accidentally launch a duplicate run and burn credits twice. Language expansion runs in translate mode for direct translations or localize mode for locale-specific wording across country-language pairs, with a sample_rate between 0.01 and 1 controlling how much of the dataset gets expanded.
The bigger play here is the loop. Dataset IDs feed directly into autoscientist.create, Adaption’s training-side counterpart that co-optimizes the data and the training recipe against your objective. Adaption claims AutoScientist beats training configured by its own research staff by an average of 35%, with win rates moving from 48% to 64% across in-house evaluations in eight verticals. Those numbers cover datasets from 5,000 to 100,000 rows. The pitch is an intent-to-trained-model pipeline with almost no human touching the intermediate steps — ambitious, and worth watching to see if those win rates hold up outside Adaption’s own eval harness.
