Adaption Labs Introduces ‘Invent a Dataset’: Training Data Generated From a Task Description, Not a Seed Corpus
Adaption Labs released the “Invent a Dataset” feature this week, which allows for the direct generation of structured training datasets based on behavioral descriptions, without the need for pre-set corpora or annotation guidelines. This feature is now available in the Adaption App, Python SDK, and REST API. The generated data is delivered in formats such as JSONL and CSV, with support for multiple languages and regional extensions. Its core mechanism involves calling the `datasets.invent` interface to automatically generate instruction pairs or preference pairs based on domain-specific codes (such as medical symptom diagnosis), and it supports credit estimation and idempotent retries. This feature aims to address the issue of limited model quality due to time-consuming data cleaning in traditional workflows, and provides a new solution for the difficult task of signal extraction in proprietary scenarios. As part of the closed loop, the generated datasets can be directly input into the `autoscientist.create` interface, allowing AutoScientist (launched in May 2026) to collaboratively optimize training recipes…