Adaption Labs has introduced "Invent a Dataset," a feature that generates structured, training-ready datasets directly from natural language task descriptions without requiring a seed corpus, predefined schema, or labeling guide. The tool is available via the Adaption app, Python SDK, and REST API, allowing users to download generated rows in JSONL, JSON, CSV, or Parquet formats.

  • A single call to `datasets.invent` initiates async generation, controlled by domain codes (e.g., medical) and subdomains.
  • Output formats include `instruction_dataset` for supervised fine-tuning and `preference_pairs` for preference-based training like DPO.
  • The API supports language expansion via `translate` or `localize` modes, with a configurable sample rate between 0.01 and 1.
  • Production parameters include an `estimate` mode for credit checking, a 10,000-character prompt limit, and idempotency keys for safe retries.
  • Generated dataset IDs feed directly into AutoScientist, which co-optimizes data and training recipes against user objectives.

Adaption reports that using AutoScientist with this workflow beats manually configured training by an average of 35%, increasing win rates from 48% to 64% across eight verticals on architectures offered by Together AI.