- Published on
Adaption Labs Introduces ‘Invent a Dataset’: Training Data Generated From a Task Description, Not a Seed Corpus
- Authors
- Name
- DeepHorus
Adaption Labs has launched a groundbreaking tool called "Invent a Dataset," which generates structured, ready-to-use training data directly from a text description of a desired model behavior. This innovative approach eliminates the need for teams to start with an existing seed corpus, predefined schema, or manual labeling guide, addressing a significant bottleneck in AI development pipelines.
The new feature is available immediately within the Adaption app, Python SDK, and REST API. Generated data can be downloaded in multiple portable formats like JSONL and CSV, ensuring users own the artifact and can use it with any training framework. However, all generation occurs on Adaption's hosted platform and requires credits, with no documented self-hosted option.
Adaption Labs argues that traditional dataset workflows, which begin with existing data, often limit model quality because available data rarely aligns perfectly with specialized tasks. The company positions "Invent a Dataset" as a solution that starts at the behavior definition level, going beyond conventional synthetic data tools that still require human-defined schemas and strategies.
The technical implementation involves a straightforward API call using domain codes, such as "medical" or subdomains like "medical.symptoms_diagnosis," to guide generation. Users can choose between two output formats: instruction datasets for supervised fine-tuning or preference pairs for reinforcement learning from human feedback (RLHF). Critical parameters include cost estimation, detailed prompting, and idempotency keys for reliable production use.
The tool also features built-in language expansion, offering both translation and localization modes to create region-specific variants, with billing based on the final expanded row count. This integration forms the first half of a "zero-data loop" that connects directly to Adaption's AutoScientist platform, launched in May 2026, for automated model training optimization.
In internal evaluations, the combined system reportedly outperforms configurations set by Adaption's own research staff, achieving a 35% improvement on average. This breakthrough has significant implications for accelerating the development of AI models for proprietary and specialized domains where relevant data is often locked in unstructured formats within internal systems.
