Adaption Labs has introduced a capability that challenges a foundational assumption in machine learning: that high-quality training data must originate from an existing corpus. With the launch of Invent a Dataset this week, the company is offering a workflow where the starting point is not a collection of files, but a sentence describing the behavior you want a model to exhibit. The feature generates a structured, training-ready dataset directly from that description, bypassing the traditional need for a seed corpus, a predefined schema, or a labeling guide.
The release arrives at a moment when the bottleneck for specialized model development has shifted from architecture to data. Teams building proprietary AI systems routinely spend weeks or months gathering, cleaning, and annotating data to approximate a target task. Invent a Dataset proposes to compress that pipeline into a single API call. The question is whether the output is genuinely deployable for production training, or whether this is a clever demo with sharp limitations.
Invent a Dataset: How the API Eliminates the Seed Corpus
The mechanics of the tool are documented in concrete terms on Adaption’s platform. A single call to datasets.inventcodecodecode creates the dataset and initiates generation, returning immediately with a status of runningcodecodecode. The developer then polls datasets.getcodecodecode until the status reads either succeededcodecodecode or failedcodecodecode, at which point the rows can be downloaded. This asynchronous pattern is standard for long-running generation tasks and makes the feature integrable into automated ML pipelines.
The primary control mechanism is the domain code. Rather than guessing or hardcoding values, developers fetch current codes using datasets.invent_domainscodecodecode. Those codes are then passed as arguments. For example, a developer working on a medical application would pass medicalcodecodecode and could optionally narrow further with qualified subdomain codes such as medical.symptoms_diagnosiscodecodecode. At least one domain or subdomain is required; passing a domain without subdomains draws from its full scope. Multiple domains can contribute to the same generation run, which is useful for tasks that sit at the intersection of fields, such as medical-legal compliance.
Two output formats are supported. The default, instruction_datasetcodecodecode, produces prompt-completion pairs suitable for supervised fine-tuning. The alternative, preference_pairscodecodecode, produces chosen and rejected completions for preference-based training methods such as Direct Preference Optimization (DPO). This dual-format support signals that Adaption is targeting both the SFT and RLHF/DPO training paradigms that dominate current LLM development.
Three additional parameters matter for production use. estimate=Truecodecodecode prices the exact request and returns estimated versus available credits without creating or charging anything, giving teams a budget check before committing compute. The promptcodecodecode parameter accepts up to 10,000 characters to steer what the rows are actually about, offering a fine-grained control surface beyond the domain code. And idempotency_keycodecodecode, which accepts up to 255 characters, makes network retries safe by returning the original dataset instead of launching a second run. Row counts are subject to a per-launch limit set by the user’s plan.
The Problem Being Targeted: Data That Already Exists
Adaption’s argument, articulated in the launch post, is that most existing dataset workflows are fundamentally constrained by the data they start with. Teams begin with a corpus they already have, then spend weeks labeling, filtering, and reshaping it to approximate the target task. This approach caps model quality at how closely the available data matches the intended behavior. For proprietary and specialized tasks, the relevant signal usually sits in internal systems, unstructured text, or workflow logs. It rarely converts cleanly into a focused training set.
The research team draws a clear distinction between Invent a Dataset and existing synthetic-data tooling. Those tools automate generation after a human has already defined the schema, task distribution, and generation strategy. The human does the hard conceptual work; the tool merely executes it faster. Invent a Dataset starts one level earlier, at the behavior itself. The developer describes what they want the model to do, and the system figures out the schema, the distribution, and the generation strategy. This is a meaningful inversion of the typical synthetic data workflow.
It is also, necessarily, a bet on the quality of the generating model. If the underlying model that powers Invent a Dataset does not understand the domain deeply enough, the generated data will contain systematic errors, hallucinations, or superficial patterns that do not generalize. Adaption does not disclose which model powers the feature, but the in-house evaluation figures — a 35% average gain over researcher-configured training — imply a level of confidence that will need to be validated by external practitioners.
Is Invent a Dataset Deployable for Production Training?
The short answer is yes, with one significant caveat. The feature is live now in the Adaption application and available through the Python SDK and REST API. Generated rows download as JSONL, JSON, CSV, or Parquet, which means the artifact is a portable file that the user owns and can train on anywhere. There is no vendor lock-in at the data level, which is a crucial design decision for trust and adoption.
The caveat concerns where generation runs. All generation executes on Adaption’s hosted platform and consumes credits. There is no documented self-hosted generation path. For teams with strict data residency requirements, or for those who prefer to keep generation within their own VPC for security reasons, this may be a blocker. Adaption documentation does not mention any on-premises or air-gapped deployment options, and the absence suggests that on-premises deployment is not a near-term priority.
For teams that can work within the hosted model, the workflow is straightforward. The dataset ID from a successful inventcodecodecode call can be passed directly to autoscientist.createcodecodecode, which co-optimizes both the data and the training recipe against the user’s objective. This closes a loop from intent to trained model.
What Is the Zero-Data Loop and How Does AutoScientist Fit?
Invent a Dataset is the first half of a larger system Adaption calls the zero-data loop. The dataset ID passes directly to AutoScientist, the training-side counterpart that launched in May 2026. AutoScientist co-optimizes the data and the training recipe against the user’s objective, meaning it can decide to drop certain rows, reweight others, or adjust hyperparameters as it iterates.
Adaption reports that AutoScientist beats training configured by its own research staff by an average of 35%. Win rates moved from 48% to 64%. These figures come from in-house domain-specialized evaluations across eight verticals. Dataset sizes ranged from 5,000 to 100,000 rows, on architectures offered for fine-tuning by Together AI. The partnership with Together AI is notable; Together AI provides the compute infrastructure for the fine-tuning runs, and the integration means users can go from dataset generation to trained model without leaving the Adaption ecosystem.
The reported 35% improvement is striking, but it is also an internal benchmark. Adaption’s researchers may have been deliberately suboptimal in their manual configurations to make the automated system look better. A more convincing test would be a blind comparison against configurations from an independent team. Still, the direction of the result aligns with what automated hyperparameter and data optimization systems have shown in other domains: systematic search often outperforms human intuition when the search space is large enough.
Language and Locale Expansion: Translate vs. Localize
The language_expansioncodecodecode feature adds another dimension of utility for multilingual deployments. It runs in two modes. translatecodecodecode produces a new row variant for each target language. localizecodecodecode produces a variant for each country-language pair, using locale-specific wording rather than direct translation. For example, localizing for Spanish would produce different variants for Spain, Mexico, and Argentina, each using regionally appropriate terminology and phrasing.
A sample_ratecodecodecode between 0.01 and 1 controls what fraction of invented rows gets expanded. This is a practical design: not every row needs to be multilingual, which would inflate costs and potentially degrade training signal for the original language. A 0.25 sample rate, for instance, means 25% of the generated rows will have language variants, giving coverage without overwhelming the dataset. Credits are billed on the expanded output row count, not the original, so users pay only for the rows they actually use.
The feature returns a 400 error for unsupported language codes, including a sample of valid values in the error message. This is a small but important detail for API usability; it means developers can debug language selection without consulting external documentation.
What Are the Practical Implications for AI Teams?
For teams building specialized models in domains like medicine, law, finance, or technical customer support, Invent a Dataset offers a path that bypasses the most painful part of the ML lifecycle: data curation. The question is not whether the generated data is perfect — it will not be — but whether it is good enough to serve as a starting point that can be refined through human review and iterative training.
The cost model is a critical factor. Adaption charges in credits, not rows, and the estimate=Truecodecodecode flag allows teams to price a request before committing. For small experiments with a few hundred rows, the cost is likely negligible. For full-scale training runs with 50,000 or 100,000 rows across multiple domains and languages, the cost could become significant. Teams should run estimates early and factor the credit cost into their budgeting.
The absence of a self-hosted option is the most obvious gap. For regulated industries — healthcare, finance, defense — sending a description of the desired behavior to a cloud API and receiving training data back may violate data governance policies. The description itself may constitute sensitive intellectual property. Adaption will need to address this with either a self-hosted deployment option or a sufficiently compelling security architecture if it wants to win those enterprise accounts.
Teams in less regulated environments, or those willing to work within Adaption’s hosted model, gain a genuine productivity advantage. The time from having an idea for a model behavior to having a training dataset shrinks from weeks to minutes. The cost of experimentation falls, which means teams can try more approaches and discard more failures before committing to a final training run.
Invent a Dataset is not a replacement for human judgment in data quality. It does not understand your edge cases, your distributional shifts, or the specific failure modes of your application. What it does is compress the distance between intention and material. That compression has real value, and it is the kind of tool that will quietly reshape how teams think about data generation in the coming years. The market will decide whether the hosted-only model is a fatal limitation or a temporary constraint that Adaption resolves in a future release. For now, the tool is live, it works, and it gives teams a new way to think about where training data comes from.
