Invent a Dataset: Data Generation for Custom Models

Today we're introducing Invent a Dataset, a new way to generate training data for custom AI models.

Instead of starting with available data, you describe what you want a model to learn. Invent a Dataset turns that objective into a structured, training-ready dataset.

This changes the starting point of dataset creation from collection to specification.

Building datasets is still one of the most manual and brittle parts of AI development.

Well-resourced teams typically start by gathering available data and then rework it to approximate the desired task. Significant time is invested in labeling, filtering, and contorting that data to fit the training objective. The result is often a dataset that only approximates the desired behavior, and a model trained around the limitations of the available data.

This creates a structural limitation. Model quality is limited by how closely the available data matches the intended behavior.

For specialized and proprietary tasks, the data you have rarely matches the behavior you want a model to learn. The relevant signals may be buried in internal systems, unstructured text, or domain-specific workflows, making them difficult to turn into a focused training dataset.

You describe the behavior you want a model to learn, such as summarizing complex business documents according to your organization’s standards, classifying customer requests across multiple departments, or following detailed policies that vary by region, product, or workflow.

Invent a Dataset interprets the objective, defines the dataset structure required to represent it, and generates the corresponding training examples.

Dataset size and scope are defined within the same flow. You do not need to arrive with a seed corpus, predefined schema, or detailed labeling guide.

Most synthetic-data tools automate generation only after someone has defined the schema, task distribution, and generation strategy. Invent a Dataset starts one level earlier: with the behavior the model should learn.

Invent a Dataset is the first half of a larger loop. Paired with AutoScientist, it turns an objective into an optimized model with no dataset to start from.

You describe the behavior you want, and Invent a Dataset generates the training signal for it. AutoScientist then co-optimizes the data and training recipe on your objective. On average, AutoScientist outperforms human-configured training by 35%.

What used to require a labeled dataset and weeks of training and tuning now starts with a description and ends with an owned model adapted to your goal.

This is what the Adaption platform is built for: Adaptive Data shapes the inputs and AutoScientist shapes the model. Intelligence shouldn't be fixed after training or improved only through slow, expensive retraining cycles. It should adapt to new objectives, new domains, and the real world.

Invent a Dataset is a step toward AI that doesn't ask you to adapt to the model, but lets you build AI that adapts to you.

Join our Discord to share feedback, ask questions, and see what others are building.