Choose data around the capability you need.
The catalog covers agentic tool use, knowledge work, multimodal reasoning, browser and computer use, and human preference. Review the task types and specifications to find a collection that fits your model.
Request a sample to inspect the inputs, context, labels, and review criteria. We can discuss where an existing collection fits and which gaps call for custom work.
Evaluate fit before choosing a collection
Compare task coverage, input format, context, and review standards with your application. Check how ambiguous examples are handled and what accompanies each label.
- Match the collection to the target behavior.
- Inspect format, context, and labeling definitions.
- Review representative samples before setting scope.
Plan the relationship to your evaluation
Track which examples are used for training and which are reserved for evaluation. Scope custom tasks for the gaps the existing collection leaves.
- Separate training and held-out evaluation uses.
- Identify domain or workflow coverage gaps.
- Confirm access, delivery, and intended use with the team.
What a program can include
Scope can include:
- Collection specifications and representative samples on request
- The context and annotations defined for the chosen dataset
- Agreed access and delivery arrangements
- A discussion of custom coverage where needed
Illustrative task / Not a production record
Make the objective concrete.
- Objective
- Select a dataset for an agent that uses several APIs.
- Context
- The intended workflows, available schemas, evaluation goals, and catalog specifications.
- Review
- Check whether samples represent the required tools, sequencing, recovery behavior, and outcomes.
Common questions
How do we request dataset samples?
Use the dataset catalog to choose the relevant collection, then submit the access request with your work email and company. The team can discuss samples, fit, and delivery requirements.
Should we choose existing data or a custom program?
Existing collections can provide a starting point when the task and standards align. Custom work is useful when your domain, environment, or failure cases require coverage that the catalog does not address.