Interpret eval results

View results in the UI

Open the Braintrust UI

Your eval will return a link to the corresponding results in Braintrust's UI. Open the link. You will land on a detailed view of the eval run that you selected.

You will see:

  • Diff mode toggle - Allows you to compare eval runs to each other. If you click the toggle, you will see the results of your current eval compared to the results of the baseline.
  • Filter bar - Allows you to focus in on a subset of test cases. You can filter by typing natural language or BTQL.
  • Summary panel - (On the right.) Shows trends across your scores, so you can home in on problematic areas.
  • Table - Shows the data for every test case in your eval run.

One eval run

Find a pattern to investigate

To find test cases to focus on, we recommend using the summary panel. Results in the panel are ordered by regressions count. This allows you to see the scorers with the biggest issues. You can also change the grouping to see summaries across any dimension in your metadata. For example, if you use separate datasets for distinct types of usecases, you can group by dataset to see which usecases are having the biggest issues. To get to an interesting subset of your test cases, click any of the filters in the summary panel.

Now that you've narrowed your test cases, you can view a test case in detail by clicking a row.

Examine the trace view

This will open the trace view. Here you can see all of the data for the trace for this test case, including input, output, metadata, and metrics for each span inside the trace.

Look at the scores and the output and decide whether the scores seem "right". Do good scores correspond to a good output? If not, you'll want to improve your evals by updating scorers or test cases.

Trace view

How metrics are calculated

Along with the scores you track, Braintrust tracks a number of metrics about your LLM calls that help you assess and understand performance.

Summary view

For example, if you're trying to figure out why the average duration increased substantially when you change a model, it's useful to look at both duration and token metrics to diagnose the underlying issue.

Wherever possible, metrics are computed on the task subspan, so that LLM-as-a-judge calls are excluded. Specifically:

  • Duration is the duration of the "task" span.
  • Prompt tokens, Completion tokens, Total tokens, LLM duration, and Estimated cost are averaged over every span that is not marked with span_attributes.purpose = "scorer", which is set automatically in autoevals.

If you are using the logging SDK, or API, you will need to follow these conventions to ensure that metrics are computed correctly.

To compute LLM metrics (like token counts), make sure you wrap your LLM cals.

Diff mode

When you run multiple experiments, Braintrust will automatically compare the results of experiments to each other. This allows you to quickly see which test cases improved or regressed across experiments.

Diff mode

How rows are matched

By default, Braintrust considers two test cases to be the same if they have the same input field. This is used both to match test cases across experiments and to bucket equivalent cases together in a trial.

Customizing the comparison key

However, sometimes your input may include additional data, and you need to use a different expression to match test cases. You can configure the comparison key in your project's Configuration page, in the "Advanced" section.

Comparison key

Aggregate (weighted) scores

It's often useful to compute many scores in your experiments (we have seen people compute hundreds!), but when reporting on an experiment, or comparing experiments over time, it's often useful to have a single score that represents the experiment as a whole.

Braintrust allows you to do this with aggregate scores, which are formulas that combine multiple scores. To create an aggregate score, go to your project's Configuration page, and click "Add aggregate score".

Add aggregate score

Braintrust currently supports two types of aggregate scores:

  • Weighted average - A weighted average of selected scores.
  • Minimum - The minimum value among the selected scores.
  • Maximum - The maximum value among the selected scores.

Analyze across experiments

Braintrust allows you to analyze data across experiments to, for example, compare the performance of different models.

Bar chart

On the Experiments page, you can select the fields you want to group by to create charts:

Group by dataset

Scatter plot

Select a metric on the x-axis to construct a scatter plot. Here's an example comparing the relationship between accuracy and duration.

Score scatterplot

Export experiments

UI

To export an experiment's results, click on the three vertical dots in the upper right-hand corner of the UI. You can export as CSV or JSON.

Export experiments

API

To fetch the events in an experiment via the API, see Fetch experiment (POST form) or Fetch experiment (GET form).

SDK

If you need to access the data from a previous experiment, you can pass the open flag into init() and then just iterate through the experiment object:

import { init } from "braintrust";
 
async function openExperiment() {
  const experiment = init(
    "Say Hi Bot", // Replace with your project name
    {
      experiment: "my-experiment", // Replace with your experiment name
      open: true,
    },
  );
  for await (const testCase of experiment) {
    console.log(testCase);
  }
}

You can use the the asDataset()/as_dataset() function to automatically convert the experiment into the same fields you'd use in a dataset (input, expected, and metadata).

import { init } from "braintrust";
 
async function openExperiment() {
  const experiment = init(
    "Say Hi Bot", // Replace with your project name
    {
      experiment: "my-experiment", // Replace with your experiment name
      open: true,
    },
  );
 
  for await (const testCase of experiment.asDataset()) {
    console.log(testCase);
  }
}

For a more advanced overview of how to reuse experiments as datasets, see Hill climbing.