Understanding Datasets

How datasets work in Future AGI: structure, column types, creation methods, and lifecycle.

About

A dataset in Future AGI is a table of structured data. Each row is one example (e.g. a user query and its expected answer). Each column is an attribute (e.g. “input”, “expected_output”, “model_response”, “score”). Datasets are the foundation for running prompts, evaluations, experiments, and optimizations.

Here’s what a simple dataset looks like:

inputexpected_outputmodel_responseis_correct
What is the capital of France?ParisParistrue
Who wrote Hamlet?ShakespeareWilliam Shakespearetrue
What is 2+2?4The answer is 4true

The first two columns (input, expected_output) are static columns that you add manually. The last two (model_response, is_correct) are dynamic columns generated by running a prompt and an evaluation against each row.


Structure

Every dataset has three core components:

  • Rows: Each row is one data point or test case. You can add rows manually, import from files, generate them synthetically, or pull them from production traces.
  • Columns: Each column defines an attribute. Columns have a name, a data type (text, number, boolean, JSON, etc.), and are either static (you provide the data) or dynamic (the platform generates it).
  • Metadata: Each dataset has a name, description, and organization-level permissions that control who can view and edit it.

How to Create a Dataset

There are several ways to get data into a dataset:

  • Manual creation: Define the structure and add rows through the UI or SDK. Learn more
  • File import: Upload CSV, Excel, JSON, or JSONL files. Learn more
  • Synthetic generation: Describe the schema and let the platform generate realistic test data. Learn more
  • From HuggingFace: Import existing datasets from HuggingFace directly. Learn more
  • From production traces: Convert observed production data from the Observe module into datasets for regression testing. Learn more

Dataset Lifecycle

1. Create

Start with a schema (columns and types) and populate it with data using any of the methods above.

2. Enrich

Add more columns to your dataset over time:

  • Run prompts: Send each row through an LLM and store the responses as a new column. Learn more
  • Run evaluations: Score model outputs using 70+ built-in metrics. Results are stored as new columns. Learn more
  • Add annotations: Manually label rows with custom tags and scores. Future AGI also supports auto-annotations that learn from your labels. Learn more

3. Experiment

Use the same dataset to compare different prompts, models, or configurations side by side. Each experiment run adds new columns so you can see results next to each other. Learn more

4. Maintain

Datasets evolve over time. You can:

  • Add or remove columns without disrupting existing data
  • Add new rows as you discover edge cases
  • Archive or delete old datasets to keep your workspace clean

Next Steps

Was this page helpful?

Questions & Discussion