Adaption Labs Introduces 'Invent a Dataset' for AI Training
Adaption Labs launches Invent a Dataset via app, Python SDK, and REST API to generate AI training data directly from task descriptions.
Stock photo for illustration only, not from the actual event
- Invent a Dataset is now live through the Adaption app, Python SDK, and REST API.
- Supports two output formats: instruction_dataset and preference_pairs.
- Features language_expansion with translate and localize modes using a sample_rate from 0.01 to 1.
- Integrates directly with AutoScientist, launched in May 2026, to co-optimize data and training recipes.
Adaption Labs has introduced a new tool named Invent a Dataset, shifting how artificial intelligence training data is prepared. The system generates training datasets directly from task descriptions rather than relying on a pre-existing seed corpus, addressing limitations in current data workflows where available information rarely fits specialized tasks.
Traditional dataset workflows typically begin with existing data, requiring weeks of labeling, filtering, and reshaping. Adaption argues that this approach caps model quality based on how closely available data matches intended behaviors. For proprietary and specialized tasks, relevant signals often reside in internal systems, unstructured text, or workflow logs that do not convert cleanly into focused training sets.
Invent a Dataset starts one level earlier at the behavior itself. A single call to datasets.invent creates the dataset and initiates generation, returning a running status immediately. Users then poll datasets.get until the status reads succeeded or failed before downloading the resulting rows.
Stock photo for illustration only, not from the actual event
Domain codes serve as the primary control mechanism, fetched via datasets.invent_domains rather than hardcoded. Users can pass values like medical, optionally narrowed by qualified subdomain codes such as medical.symptoms_diagnosis. Two output formats are supported: instruction_dataset for supervised fine-tuning and preference_pairs for preference-based training like DPO.
For production use, three parameters are available: estimate=True for exact pricing checks without charges, prompt accepting up to 10,000 characters to steer row generation, and idempotency_key accepting up to 255 characters for safe network retries. Additionally, language_expansion operates in translate and localize modes with a sample_rate between 0.01 and 1.
Invent a Dataset represents a crucial step in resolving the data bottleneck in AI training, which traditionally demands significant time and labeling effort. Generating data directly from task specifications allows engineering teams to iterate faster without waiting for manual collection pipelines.
Invent a Dataset forms the first half of a loop, passing dataset IDs directly to autoscientist.create to co-optimize data and training recipes against specific objectives. AutoScientist launched in May 2026, and in-house evaluations across eight verticals showed it outperforms training configured by research staff by an average of 35% on Together AI architectures.
Source: MarkTechPost
Found something wrong in this article? Report an issue with this article
Comments
Leave a Comment