Editorial illustration for Adaption Labs’ ‘Invent a Dataset’ Generates Training Data From Task Descriptions
Adaption Labs Generates Training Data From Task Descriptions
Adaption Labs launched a feature this week called Invent a Dataset, and it skips a step that most synthetic-data tools treat as fixed: the schema. Instead of feeding the system a seed corpus or a labeling guide, a user hands it a description of the behavior they want a model to learn. The system builds a structured, training-ready dataset from that description alone.
The company is positioning this against the standard workflow, where teams start with data that already exists, then spend weeks labeling and reshaping it to fit the target task. That approach caps model quality at how closely the available data resembles the intended behavior, according to Adaption. For specialized or proprietary tasks, the useful signal usually lives in internal systems or workflow logs that don't convert into a clean training set without a lot of manual work.
The tool is live now in the Adaption app, plus a Python SDK and REST API, with output downloadable as JSONL, JSON, CSV, or Parquet. Generation runs on Adaption's hosted platform and draws on credits. No self-hosted option is documented yet.
The research team also draws a line against existing synthetic-data tooling. Those tools automate generation after a human has already defined the schema, task distribution, and generation strategy. Invent a Dataset starts one level earlier, at the behavior itself.
Why this matters
For teams building on top of foundation models, the bottleneck has rarely been model quality. It's been the grunt work of assembling a labeled corpus before you can even test whether fine-tuning helps. Invent a Dataset removes that first step by letting a task description stand in for a seed corpus, which changes the economics of experimentation. You can test a training idea in an afternoon instead of a sprint.
The billing detail matters more than it looks. Charging on expanded row count rather than invented rows means the sample_rate knob isn't just a quality lever, it's a cost lever, and teams will need to treat it that way from day one rather than discovering it on an invoice. The 400 error on unsupported codes is a small thing, but it signals Adaption Labs built this for production use, not just demos.
We'd still want to see how these invented datasets hold up against real-world edge cases before trusting them for anything customer-facing. Synthetic data that looks clean in a sample can still encode blind spots nobody wrote down.
Related Reading
Research & Benchmarks Google's FACTS benchmark shows 70% factuality ceiling across four tests
Research & Benchmarks Databricks finds multi-step agents beat single-turn RAG by 21% to 38% on STaRK
Research & Benchmarks Nvidia's DLSS 4.5 beta adds 6x Multi Frame Generation for RTX 50 GPUs
From Archives OpenAI Agents Accessed German Wiki to Share Sandbox Exploits, Logs Show
From Archives GPT-6 Astra solves two open Erdős problems with USD 300 budget
Common Questions Answered
How does Adaption Labs' Invent a Dataset differ from traditional synthetic-data tools?
Invent a Dataset skips the schema definition step that most synthetic-data tools require, allowing users to input only a task description instead of a seed corpus or labeling guide. Traditional tools automate generation after humans have already defined the schema, task distribution, and generation strategy, whereas Invent a Dataset starts at the behavior level itself and builds the structured dataset from that description alone.
What is the main bottleneck that Invent a Dataset addresses for teams building on foundation models?
The primary bottleneck has been the time-consuming grunt work of assembling a labeled corpus before testing whether fine-tuning helps. Invent a Dataset removes this initial step by letting a task description replace the need for a seed corpus, enabling teams to test training ideas in an afternoon instead of spending an entire sprint on data preparation.
What input does a user need to provide to Invent a Dataset to generate training data?
Users only need to provide a description of the behavior they want a model to learn, rather than pre-existing data, labeling guides, or predefined schemas. The system then builds a structured, training-ready dataset directly from that task description alone.
How does Invent a Dataset change the economics of machine learning experimentation?
By eliminating the need for manual data labeling and schema definition upfront, Invent a Dataset significantly reduces the time and resources required to test training ideas. This allows teams to move from concept to testing in a single afternoon rather than requiring weeks of data preparation work.
Further Reading
- Invent a Dataset: Data Generation for Custom Models - Adaption Labs
- Invent a Dataset | Create Training Datasets from Intent - Adaption Labs
- Adaption Labs introduces Invent a Dataset - 鉅亨網
- Introducing Invent a Dataset – Adaption is introducing Invent a Dataset, a new way to generate training data for custom AI models - Radical Data Science
- Adaptive Data - Adaption Labs Docs
