Adaption launches a dataset generator that starts with instructions, not data

Why it matters

Invent a Dataset could let teams build owned, task-specific models before collecting a corpus, making synthetic data the entry point to Adaption's full training stack.

Sara Hooker (@sarahookr) and Sudip Roy launched Invent a Dataset on September 3rd, giving Adaption users a way to generate training data by describing what they want a custom AI model to learn.

The founders are pushing dataset creation one step upstream. Instead of asking users to supply a seed corpus, define a schema, or write labeling instructions, Invent a Dataset interprets a plain-English objective, chooses a structure, and generates training examples at the requested size and scope. The output is delivered asynchronously as a training-ready dataset, according to Adaption.

Hooker and Roy built Adaption around the view that AI systems should be cheaper to modify and more responsive to narrow, changing requirements. Hooker previously led research at Cohere and worked at Google Brain and Google DeepMind. Roy led inference and platform work at Cohere after working on AI infrastructure at Google, including Pathways. Their backgrounds explain the product's emphasis: Adaption is treating custom training as an engineering workflow that can be automated, rather than a research project that starts with months of data preparation.

Adaption moves upstream

In its launch post, Adaption gives examples including a model that summarizes business documents according to an organization's standards, routes customer requests across departments, or follows policies that vary by product and region. Its product page also shows proposed datasets for hotel reviews, customer support conversations, Vietnamese real estate listings, and medical question-answer pairs.

Most synthetic-data products still require someone to decide what each row should contain, how tasks should be distributed, and which prompts or labels should be generated. Adaption says Invent a Dataset handles those decisions from the behavioral specification itself. That makes the prompt the starting artifact for model development, replacing the usual pile of documents, tables, and annotation rules.

The approach is useful when the target behavior is clear and the underlying examples are scarce, fragmented, confidential, or expensive to label. It also moves more responsibility into the generation system. A plausible-looking dataset can still teach the wrong behavior if the generator misreads a policy, underrepresents an edge case, or reproduces errors across thousands of examples. That risk becomes sharper in the regulated-industry use cases Adaption is courting, where fluent medical or financial answers are not necessarily correct ones.

Invent a Dataset is designed to feed directly into AutoScientist, Adaption's system for automating model selection and fine-tuning. AutoScientist searches training recipes, adjusts the data, trains candidate models, and evaluates them against the user's objective. The combined pitch is an end-to-end path from a written specification to downloadable model weights, without an existing dataset.

Adaption says AutoScientist beat configurations chosen by its own AI researchers by an average of 35% across internal tests. The underlying results covered datasets ranging from 5,000 to 100,000 examples, several model architectures available through Together AI, and eight verticals. Adaption reported that aggregate win rates rose from 48% with researcher-selected configurations to 64% with AutoScientist. Those are company-run, domain-specific evaluations, and Invent a Dataset introduces another variable: whether generated examples accurately encode the requested behavior before training begins.

A larger bet on owned models

The release fills a missing input in Adaption's product sequence. Adaptive Data, introduced in February, reshapes and expands existing datasets. AutoScientist, released in May, automates the training loop. Invent a Dataset lets that loop begin without source material, extending Adaption's reach to teams that know the behavior they need but have not assembled examples for it.

That is also a distribution play. Custom-model platforms typically enter after a customer has collected data and committed engineering resources to fine-tuning. Invent a Dataset gives Adaption a reason to enter at the idea stage, then keep the workload inside its platform through data generation, training, evaluation, and deployment.

The product arrives seven months after Adaption raised a $50 million seed round led by Emergence Capital. Mozilla Ventures, Fifty Years, Threshold Ventures, Alpha Intelligence Capital, E14 Fund, and Neo also participated. The unusually large seed financed Hooker and Roy's challenge to the industry's scale-first model: use automated data and training systems to produce smaller, task-specific models instead of sending every workload to a general-purpose frontier API.

Invent a Dataset turns that thesis into a simpler interface. The customer supplies an outcome. Adaption supplies the examples and training process. The quality of the resulting model will depend on whether the generated dataset captures the difficult cases that a short specification leaves unsaid.