Adaption labs launches tool to generate training data from task descriptions

Adaption Labs has launched a new feature called 'Invent a Dataset,' which allows users to generate structured, training-ready datasets directly from a task description, without requiring a seed corpus, predefined schema, or labeling guide. The tool is currently available through the Adaption app, Python SDK, and REST API, with generated data exported in formats like JSONL, JSON, CSV, or Parquet. The process runs on Adaption’s hosted platform and consumes credits, with no self-hosted generation path documented. The tool aims to address limitations in traditional dataset workflows, where teams often spend weeks labeling and reshaping existing data to match target tasks. Adaption argues that this approach limits model quality, as relevant signals are often found in internal systems, unstructured text, or workflow logs. Invent a Dataset starts earlier in the process, focusing on the behavior itself rather than predefined schemas. It supports two output formats for training: instruction_dataset for supervised fine-tuning and preference_pairs for preference-based training like DPO. Adaption also introduced AutoScientist, launched in May 2026, which co-optimizes data and training recipes against user objectives. Internal evaluations show AutoScientist outperforms manually configured training setups by an average of 35%, with win rates increasing from 48% to 64% across eight verticals. Dataset sizes ranged from 5,000 to 100,000 rows, using architectures from Together AI.