How to map ExperimentEvaluationRun to ExperimentRun

[andstor]

I need to export all the data from an experiment. The goal is to have a df with all the experiment data, including evaluation annotations.

Using get_experiment from phoenix.client.experiments gets me nearly there. However, I can not find any way to map an ExperimentEvaluationRun to an ExperimentRun.

[dosubot[bot]]

You can map an ExperimentEvaluationRun to its corresponding ExperimentRun using the experiment_run_id field, which matches the id of the ExperimentRun docs. If you’re using get_experiment from phoenix.client.experiments, it returns a RanExperiment object that already aggregates both runs and evaluations.

To export all experiment data, including evaluation annotations, as a single DataFrame, use the get_evaluations() method on the RanExperiment object. This method joins evaluation runs to experiment runs using the run ID, so you get all the experiment data and evaluation results in one DataFrame source:

import phoenix as px

# Retrieve the experiment (replace with your experiment_id)
experiment = px.Client().get_experiment(experiment_id="your-experiment-id")

# Export all experiment data, including evaluation annotations
df = experiment.get_evaluations()

The resulting DataFrame will have one row per experiment run, with columns for run data (input, output, etc.) and evaluation results (score, label, explanation, etc.). The join is handled for you by matching ExperimentEvaluationRun.experiment_run_id to ExperimentRun.id, so you don’t need to do any manual mapping or merging. If you want just the run data without evaluations, use experiment.as_dataframe().

This approach works even if some runs don’t have evaluations—they’ll still appear in the DataFrame, with evaluation columns as NaN source.

To reply, just mention @dosu.

---

Share context across your team and agents. Try Dosu.

[mikeldking]

Hey @andstor there are a few ways I believe this is sorta possible today.
one is through the UI and one is through the REST api (https://arize.com/docs/phoenix/sdk-api-reference/rest-api/experiments)

Would either of these work for you? or would you prefer an explicit method in the client?

Mainly trying to understand the use-case for the experiment as a dataframe. Happy to support whatever use-case you may have.

[andstor]

Hi @mikeldking, yes, that works — thanks!
However, I’d prefer a more convenient way to retrieve the results directly from the client. My use case is that I need to calculate aggregate metrics based on the experiment run results.

Btw, it would also be great to have a way to log aggregate metrics and display them on the experiments overview page.

[mikeldking]

@andstor This makes a lot of sense to me too. What sort of aggregate metrics are you looking for? TPR, TNR, precision, recall come to mind. Any other use-cases?

[mikeldking]

Quick update on this. We'll make the annotations that are the result of annotations available in in the client. #8918

[andstor]

@mikeldking Great!

And yes, that would be some of the most common and useful metrics.

[JohannKaspar]

Same here, this feature would be very valuable for unbalanced datasets. I am not sure whether #8918 actually addresses this.