Repetitions

How to leverage repetitions to get an understanding of indeterminate LLM outputs

repetitions are available in Phoenix 11.37.0 and the clients that support them

Since LLMs are probabilistic, their synthesis can differ even when the supplied prompts are exactly the same. This can make it challenging to determine if a particular change is warranted as a single execution cannot concretely tell you whether a given change improves or degrades your task.

So what can you do when an execution can change from one run to the next? That’s where repetitions come in. Repetitions help you reduce uncertainty in systems prone to variability, notably more “agentic” systems.

Configuring Repetitions

Repetitions can be configured whenever you run an experiment via the phoenix client. The repetitions parameter determines how many times each example is used in your task. So if you have 3 examples with 2 repetitions, your task will be run 6 times and evaluated 6 times.

Python

from phoenix.client import Client
client = Client()
dataset = client.datasets.get_dataset(dataset="my-dataset")

def my_task(input):
    return f"Hello {input['name']}"

experiment = client.experiments.run_experiment(
    dataset=dataset,
    task=my_task,
    experiment_name="greeting-experiment"
    repetitions=3
)

Viewing Repetitions

If you’ve run your experiments with repetitions, you will see arrow icons at the top of each output. At the bottom you will see the average of the evaluations as well as the score for the evaluation you are looking at. You can click on the arrows to cycle through the repetions or click on the expand icon to view the full details.