Manage datasets in the console

Create a dataset

1. Log on to the AgentLoop console and select the target workspace.

2. Open the Data Center module. On the Datasets page, click Create Dataset. The creation page contains configuration areas for basic information, data source, and field definitions.

   Basic information:

   • Dataset name: Use lowercase letters, digits, and underscores for easy querying and management.

   • Description: Describe the purpose of the dataset. For example, "customer service QA evaluation set", "trace cleaning sample set", or "model alignment test set".

   Data source: Three options are available.

   • From trace data: Best suited for scenarios where agent runtime traces already exist. The system extracts fields from ingested trace data and generates a dataset, enabling automated construction of evaluation samples.

   • Upload CSV / Excel / JSONL files: Suitable for offline sample imports. After you upload local files, the system recognizes the field structure and assists with dataset creation.

   • Start from scratch: For fully custom datasets. You can manually define field names, field types, vector indexes, and Chinese language support settings.

   Field definitions: If you choose to start from scratch, configure the data structure in the field definitions section.

   Fields typically include the field name, field type, vector index configuration, and Chinese language support. Common fields include question, expected_output, and context. Enable vector indexes on text fields that require semantic retrieval.

Manage a dataset

1. Log on to the AgentLoop console and select the target workspace.

2. Open the Data Center module. On the Datasets page, click the name of the target dataset to open its details page.

3. The details page provides the following functional areas:

   • Data Details: View, query, and maintain data records. You can add records, upload data in bulk, or run query statements in the query box to verify data writes or check data distribution.

   • Field Management: View and manage the current dataset schema, including field names, types, vector indexes, and Chinese language support.

   • Annotation: Manage sample annotation and downstream evaluation workflows for the dataset.

4. On the data details page, the console provides a list view toggle, hover preview switch, query input box, and an execute query button. Use query statements to inspect data content, such as counting records or filtering samples by specific fields. Query results are displayed in a table for quick status verification.

   Write and view data

   Datasets support two common write methods: adding individual records and uploading data in bulk. For small sample sizes, click Add Record to enter data manually. For existing CSV files or batch samples, use Bulk Upload for fast imports. When a dataset is empty, the details page prompts you to add data through either method. After data is written, use the query entry point to validate results, such as checking totals, filtering by field, or inspecting specific sample content.

   Process data from traces

   For trace-sourced data, the console provides data filtering and processing entry points. You can filter trace data by agent application and expand advanced filters to further narrow the data range. Filtered data can proceed to downstream processing workflows for field extraction, sample cleaning, and generation of evaluation-ready datasets. This approach lets you accumulate inputs, outputs, contexts, and reference answers from actual agent runs into evaluation data, reducing manual curation costs and enabling datasets to continuously serve evaluation, experimentation, and model optimization.

Delete a dataset

A delete button is available on the dataset details page. Confirm the operation before proceeding.