Adaption Labs generates a training dataset from a description

Adaption Labs has launched Invent a Dataset, a service that generates instruction or preference datasets from a description without requiring an initial corpus or predefined schema.

Describe what a model should learn, choose a domain, and specify the desired number of examples. Invent a Dataset promises to handle what follows: define the dataset structure, select relevant material, and produce the examples required for training.

Introduced by Adaption Labs, the service reverses the usual starting point for a model customization project. A team no longer necessarily begins with the documents, conversations, or tables it already owns. It specifies the desired behavior, then lets the platform create a dataset around that specification.

The company cites tasks such as summarizing documents according to an organization’s standards, classifying requests sent to customer support, and following policies that vary by region or product. Other examples shown on the service page include hotel reviews, technical support conversations, Vietnamese real estate listings, and medical questions.

The approach addresses a familiar problem. Existing data was rarely created to train a model. It may contain the right information without presenting it as an instruction paired with the expected response. Turning it into training material can require cleaning, deduplication, category design, example writing, and sometimes an extensive annotation campaign.

Invent a Dataset attempts to move that work into a specification interface. The user selects at least one domain or subdomain, sets a target number of rows, and can add a free-form description of up to 10,000 characters. This text can define the subject, audience, level of detail, or expected writing style.

The word “invent” might suggest that the tool freely creates every example from a single sentence. The technical documentation adds an important qualification: the description is used for a semantic search that selects source material, then incorporated into the generation instructions.

The service can therefore operate without a customer-provided corpus, but not without data in the broader sense. It relies on material selected by the platform and on models capable of transforming it. In its launch materials, Adaption Labs does not fully disclose the composition of those sources, the models used to generate the examples, or the provenance checks applied to individual rows.

Two training structures are currently available. The default format produces pairs containing an instruction and a response. It is designed for supervised fine-tuning, commonly abbreviated as SFT.

The second generates preference pairs. Two responses are associated with the same instruction, with one marked as chosen and the other as rejected. This material can support methods such as DPO, which train a model to favor particular behaviors without requiring a numerical score for every response.

Generation runs asynchronously. A request creates the dataset and immediately returns its identifier with a `running` status. The application must then poll the service until processing is complete before downloading the result.

The requested number of rows is a target rather than a strict guarantee. The documentation warns that the delivered volume may land slightly above or below that figure. Per-run limits also apply depending on the customer’s plan.

Once processing finishes, the data can be exported as JSONL, JSON, CSV, or Parquet. Instruction datasets place the generated text in fields reserved for the enhanced prompt and completion. Fields corresponding to original examples remain empty because the rows were not directly derived from customer-supplied records.

Invent a Dataset is accessible through the Adaption Labs web application and API. Its Python SDK includes methods for retrieving the current domain list, estimating a request’s cost, starting generation, tracking its status, and downloading the completed file.

The estimate is optional but useful for larger requests. It reports the number of required credits without starting the job or charging the account. Adaption Labs still describes Invent’s pricing as provisional and does not publish a general per-row rate on the launch page. Credit consumption is calculated from the number of rows produced after any language expansion.

The platform can translate a sample of the examples into several languages or localize them for specific country-and-language combinations. Localization is intended to adjust wording for a particular market rather than merely replace words with their equivalents. Each additional variant, however, increases the billable output volume.

This feature could simplify the preparation of models for markets where suitable data is scarce. It does not remove the need for local review. A grammatically correct translation may use the wrong register, overlook an administrative practice, or treat a regional rule as universal.

The same caution applies in regulated industries. A collection of medical or legal questions may be stylistically consistent without being accurate, current, or suitable for a particular jurisdiction. Generation reduces the writing workload, but it does not transfer responsibility for validating facts and business rules to the system.

The initial description therefore becomes a critical component. If it omits a rare case, contractual exception, or ambiguous category, the resulting dataset may reproduce that omission. A system that expands a specification automatically can introduce more variety, but it cannot reliably infer requirements that the organization never expressed.

This limitation is particularly relevant to tasks described as proprietary. Without internal documents, Invent a Dataset can create examples resembling a refund process, commercial policy, or financial report. It cannot know a company’s specific rules unless they appear in the description or in material supplied to the platform.

Adaption Labs separately offers Forge for transforming existing documents and Adaptive Data for restructuring or enriching them. Invent a Dataset occupies a different position: it provides a starting point without a