> For clean Markdown of any page, append .md to the page URL.
> For a complete documentation index, see https://docs.astropods.com/llms.txt.
> For AI client integration (Claude Code, Cursor, etc.), connect to the MCP server at https://docs.astropods.com/_mcp/server.

# ast dataset

Inspect and curate the evaluation datasets in the active account. A dataset holds the traces you have curated from a deployed agent, with their evaluator values.

```bash
ast dataset <subcommand>
```

Every subcommand except `list` takes a dataset name. Run `dataset list` to see the names. The list shows datasets for currently deployed agents.

## dataset list

```bash
ast dataset list [--limit <n>] [--offset <n>] [--json]
```

Lists the account's datasets, newest first, with the blueprint and deployment each belongs to. The default `--limit` is 50 and the maximum is 100. When more datasets remain, the output prints the `--offset` for the next page.

## dataset get

```bash
ast dataset get <dataset-name> [--json]
```

Shows the dataset's item count and, for each evaluator, how many items have each value.

## dataset items

```bash
ast dataset items <dataset-name> [--limit <n>] [--offset <n>] [--json]
```

Lists a page of items with their input, expected output, and evaluator values. The default `--limit` is 50 and the maximum is 100. `--offset` must be a multiple of `--limit`. Items whose values come from an earlier evaluation set are marked `(outdated)`. Use `--json` for the full, untruncated values.

## dataset add

```bash
ast dataset add <dataset-name> --trace-id <trace-id> [--json]
```

Adds the trace to the dataset as an item, with its current evaluator values. The trace must belong to the dataset's deployment and have an input. Adding a trace that is already in the dataset fails.

## dataset edit

```bash
ast dataset edit <dataset-name> --trace-id <trace-id> \
  --set <key>=<value> [--set <key>=<value> ...] \
  [--set-string <key>=<value> ...] [--json]
```

Replaces the item's evaluator values. Evaluators you leave out are cleared. Get keys and accepted values from [`ast eval get`](/cli/eval#eval-get).

* `--set` reads `true` and `false` as booleans, numbers as numbers, and anything else as text.
* `--set-string` always reads the value as text.

An item added under an earlier evaluation set cannot be edited. Remove it and add it again.

## dataset remove

```bash
ast dataset remove <dataset-name> --trace-id <trace-id> [--confirm <trace-id>] [--json]
```

Removes the trace's item and its evaluator values from the dataset. The trace and its evaluations are kept. The command asks for confirmation. Pass `--confirm` with the trace ID to skip the prompt.

## dataset download

```bash
ast dataset download <dataset-name> [--output <path>]
```

Saves the dataset as a zip file containing one JSONL file of its items. The command writes `<dataset-name>.zip` to the current directory. `--output` (`-o`) sets a file or an existing directory, and `--output -` writes the zip to standard output.

## Next steps

* [ast eval](/cli/eval): evaluate traces and save the reviews that fill a dataset
* [ast agent](/cli/agent#agent-trace): find the trace to add