Adaption Labs Releases Dataset Generator From Task Description

Adaption Labs released its 'Invent a Dataset' feature this week, a novel tool designed to generate structured, training-ready datasets directly from a textual description of desired model behavior. This approach bypasses the traditional requirement of starting with a seed corpus, a predefined schema, or a detailed labeling guide. The feature is currently live and accessible through the Adaption app, as well as via a Python SDK and a REST API. Generated datasets can be downloaded in common formats including JSONL, JSON, CSV, and Parquet, ensuring portability and ownership for users to train models on any platform. The generation process itself is hosted on Adaption's platform and operates on a credit-based system, with no documented option for self-hosted generation.

The core problem 'Invent a Dataset' aims to solve is the inherent limitation in traditional dataset workflows, which often begin with existing data. Teams typically spend considerable time labeling, filtering, and reshaping this data to align with a target task. Adaption Labs argues that this process inherently caps model quality, as it is constrained by how closely the available data mirrors the intended behavior. For specialized or proprietary tasks, the crucial signals are often embedded within internal systems, unstructured text, or workflow logs, and these rarely translate cleanly into a focused training set without extensive manual effort. Furthermore, Adaption distinguishes its offering from existing synthetic data tools. While those tools automate generation based on human-defined schemas, task distributions, and generation strategies, 'Invent a Dataset' operates at a more fundamental level, starting directly with the desired behavior itself.

The mechanics of the API are clearly documented. A single call to the `datasets.invent` endpoint initiates dataset creation and begins the generation process, returning an immediate status of 'running'. Users then poll the `datasets.get` endpoint to monitor progress until the status indicates 'succeeded' or 'failed', at which point the generated rows can be downloaded. A key control mechanism involves 'domain codes', which users can fetch using `datasets.invent_domains` rather than hardcoding them. These codes, such as 'medical', can be optionally narrowed down with qualified subdomain codes like 'medical.symptoms_diagnosis'. At least one domain or subdomain is mandatory for a generation run, and multiple domains can be specified to contribute to a single dataset generation. This structured approach allows for granular control over the type of data generated, ensuring relevance to specific AI training objectives.