> ## Documentation Index
> Fetch the complete documentation index at: https://docs.bryel.ai/llms.txt
> Use this file to discover all available pages before exploring further.

# Datasets

> Curate production traces into supervised fine-tuning (SFT) datasets, judge their quality, and export them for training.

A dataset is a curated set of records for **supervised fine-tuning**. Each record is one agent run, reconstructed into a chat-with-tools trajectory (`{messages, tools}`) — the training target. There's no separate "expected" answer: SFT learns from the trajectory itself.

## Build a dataset

Datasets are built from your real runs. Create one, then add the traces worth training on.

<Tabs>
  <Tab title="In the app">
    <Steps>
      <Step title="Create a dataset">
        From **Datasets**, give it a name. Empty to start.
      </Step>

      <Step title="Add traces">
        Use the trace selector to filter your runs (by intent, status, cost, tools, and more), preview the matches, and add them. Each trace is reconstructed into a record with its full message + tool trajectory.
      </Step>
    </Steps>
  </Tab>

  <Tab title="From a coding agent">
    With the [remote MCP](/guides/coding-agents) and write access, an agent builds a set from your editor:

    ```
    "Using bryel: create a dataset 'good-add-pricing', add every clean
    add_pricing run from the last 30 days, and inspect the result."
    ```

    That runs `bryel_create_dataset` → `bryel_add_from_query` → `bryel_inspect_dataset`. Adding from a query is idempotent per trace, so you can widen the filter and re-run without duplicating.
  </Tab>
</Tabs>

## Judge quality before you train

A fine-tune is only as good as its data. bryel surfaces the dimensions that make or break an SFT set — in the dataset view and through `bryel_inspect_dataset`.

<ParamField path="Balance" type="intent distribution">
  How records spread across intents. One intent dominating makes the model overfit it and forget the rest.
</ParamField>

<ParamField path="Redundancy" type="duplicate clusters">
  Near-identical inputs waste capacity and bias the model. bryel groups them so you can prune to a few exemplars.
</ParamField>

<ParamField path="Trajectory health" type="looping / errored">
  Runs that looped or errored teach the model bad behavior. They're flagged so you can drop or repair them.
</ParamField>

<ParamField path="Length" type="token & step distribution">
  Records that exceed the fine-tune context window get truncated mid-trajectory; trivially short ones add no signal.
</ParamField>

In the dataset view you can filter to **duplicates** or **problematic** records and prune them in place.

## Export for training

Export a dataset to fine-tune **JSONL** — one `{messages, tools}` object per line, ready for a training run.

<Steps>
  <Step title="Export JSONL">
    From the dataset, choose **Export JSONL** — bryel streams the records to a file you can download. A [coding agent](/guides/coding-agents) can do the same with `bryel_export_dataset`, then poll `bryel_export_status` for the download URL.
  </Step>

  <Step title="Train">
    Feed the JSONL to your fine-tuning pipeline.
  </Step>
</Steps>

<Card title="Build datasets from your editor" icon="bot" href="/guides/coding-agents" horizontal>
  Connect a coding agent over the remote MCP to query, curate, and inspect datasets.
</Card>