Run offline evals on experiments

Eval-driven development

You have validated evaluators and a golden dataset. Now make them part of your development process. Every time you change a prompt, swap a model, or update your agent logic, score your experiment results against your golden dataset before deploying. Each evaluation runs in a controlled, repeatable environment so you can measure how new versions behave before exposing them to real users. The same evaluators you use for production monitoring work here. Define the criteria once, apply them everywhere. This page covers running evaluators on experiments. To create experiments and datasets, see Datasets and experiments.

Experiments tab with summary metrics chart and experiments table, row menu open with View Eval Trace and View Logs

Run on an experiment

Once you have an experiment, attach evaluators and score the results.

By Arize Skills
By Alyx
By UI
By Code

Use the arize-evaluator skill to create an eval task against an experiment via the ax CLI. Install the Arize skills plugin in your coding agent if you have not already. Then ask your agent:

“Create an eval task for my v2-prompt-test experiment using my correctness evaluator”
“Trigger an eval run on my latest experiment and wait for results”
“Score my v2-prompt-test experiment with my hallucination evaluator”

Terminal showing ax datasets create and help output, ax experiments help, and an in-progress v2-prompt-test experiment with steps to create dataset, run experiment, create evaluator, create eval task, and trigger run

The skill resolves dataset and experiment IDs, configures column mappings, and triggers the run using ax tasks create and ax tasks trigger-run.

Here’s the simplest version of an evaluation function:

def is_true(output):
    # output is the task output
    return output == True

Evaluation Inputs

The evaluator function can take the following optional arguments:

Parameter name	Description	Example
`dataset_row`	the entire row of the data, including every column as dictionary key	`def eval(dataset_row): …`
`input`	experiment run input, which is mapped to `attributes.input.value`	`def eval(input): …`
`output`	experiment run output	`def eval(output): …`
`dataset_output`	the expected output if available, mapped to `attributes.output.value`	`def eval(dataset_output): …`
`metadata`	dataset_row metadata, which is mapped to `attributes.metadata`	`def eval(metadata): …`

Evaluation Outputs

We support several types of evaluation outputs. Label must be a string. Score must range from 0.0 to 1.0. Explanation must be a string.

Evaluator Output Type	Example	How it appears in Arize
`boolean`	`True`	label = ‘True’ score = 1.0
`float`	`1.0`	score = 1.0
`string`	`”reasonable”`	label = ‘reasonable’
`tuple`	`(1.0, “my explanation notes”)`	score = 1.0 explanation = ‘my explanation notes’
`tuple`	`(“True”, 1.0, “my explanation”)`	label = ‘True’ score = 1.0 explanation = “my explanation”
`EvaluationResult`	`EvaluationResult(` `score=1,` `label=‘reasonable’, explanation=‘explanation’` `metadata=` `)`	score = 1.0 label=‘reasonable’ explanation = ‘explanation’ metadata=

To use EvaluationResult class, use the following import statement:

Version 7: from arize.experimental.datasets.experiments.types import EvaluationResult
Version 8: from arize.experiments import EvaluationResult
One of label or score must be supplied (you can’t have an evaluation with no result).

Here is an example of an evaluator which compares the output to a value in the dataset_row.

from arize.experiments import EvaluationResult
import pandas as pd

# Example dataset
inventions_dataset = pd.DataFrame({
    "attributes.input.value": ["Telephone", "Light Bulb"],
    "attributes.output.value": ["Alexander Graham Bell", "Thomas Edison"],
})

def is_correct(output, dataset_row):
    expected = dataset_row.get("attributes.output.value")
    correct = expected in output
    return EvaluationResult(
        score=int(correct),
        label="correct" if correct else "incorrect",
        explanation="Evaluator explanation here"
    )

To run the experiment:

experiment, experiment_df = client.experiments.run(
    name="basic-experiment",
    dataset=dataset_id,
    task=answer_question,
    evaluators=[is_correct],
    concurrency=10,
    exit_on_error=False,
    dry_run=False,
)

Create an LLM Evaluator

LLM evaluators utilize LLMs as judges to assess the success of your experiment. These evaluators can either use a prebuilt LLM evaluation template or be customized to suit your specific needs.Arize supports a large number of LLM evaluators out of the box with LLM Classify: Arize Templates. You can also define custom LLM evaluators. Here’s an example of a LLM evaluator that checks for correctness in the model output:

CORRECTNESS_PROMPT_TEMPLATE = """
You are given an invention (input) and an inventor (output). Determine whether the inventor correctly corresponds to the invention.

[BEGIN DATA]
[Inventor]: {invention}
[Output]: {output}
[END DATA]

Explain your reasoning step by step, then provide a single-word LABEL at the end: either "correct" or "incorrect".

Format:

EXPLANATION: Your reasoning about why the output is correct or incorrect
LABEL: "correct" or "incorrect"
************
"""

Run Evaluation

from phoenix.evals import llm_classify, OpenAIModel
from arize.experiments import EvaluationResult
import pandas as pd

def correctness_eval(output, dataset_row):
    # Get the original query topic
    invention = dataset_row.get("attributes.output.value")

    eval_df = llm_classify(
        dataframe=pd.DataFrame([{"invention": invention, "output": output}]),
        template=CORRECTNESS_PROMPT_TEMPLATE,
        model=OpenAIModel(model="gpt-4o-mini", api_key="your-openai-api-key"),
        rails=["correct", "incorrect"],
        provide_explanation=True,
    )

    # Map the eval df to EvaluationResult
    label = eval_df["label"][0]
    score = 1 if label == "correct" else 0
    explanation = eval_df["explanation"][0]

    return EvaluationResult(label=label, score=score, explanation=explanation)

In this example, correctness_eval evaluates whether the output of an experiment is correct. The llm_classify function runs the eval, and the evaluator returns an EvaluationResult that includes a score, label, and explanation.Once you define your evaluator class, you can use it in your experiment run like this:

experiment, experiment_df = client.experiments.run(
    name="test-experiment",
    dataset=dataset_id,
    task=answer_question,
    evaluators=[correctness_eval],
)

You can customize LLM evaluators to suit your experiment’s needs — update the template with your instructions and the rails with the desired output.

Create a Code Evaluator

Code evaluators are functions designed to assess the outputs of your experiments. They allow you to define specific criteria for success, which can be as simple or complex as your application requires. Code evaluators are especially useful when you need to apply tailored logic or rules to validate the output of your model.

Custom Code Evaluators

Creating a custom code evaluator is as simple as writing a Python function. By default, this function will take the output of an experiment run as its single argument. Your custom evaluator can return either a boolean or a numeric value, which will then be recorded as the evaluation score.For example, let’s say our experiment is testing a task that should output a numeric value between 1 and 100. We can create a simple evaluator function to check if the output falls within this range:

def in_bounds(output):
    return 1 <= output <= 100

By passing the in_bounds function to run_experiment, evaluations will automatically be generated for each experiment run, indicating whether the output is within the allowed range. This allows you to quickly assess the validity of your experiment’s outputs based on custom criteria.

experiment, experiment_df = client.experiments.run(
    name=experiment_name,
    dataset=dataset_id,
    task=answer_question,
    evaluators=[in_bounds],
)

Prebuilt Phoenix Code Evaluators

You can also leverage our open-source Phoenix pre-built code evaluators. Pre-built evaluators can be passed directly to the evaluators parameter when running experiments.

Use dry_run=True to test without logging results. Use concurrency=10 to speed up large runs. Start with synchronous evaluators when debugging, then switch to async for speed.

For class-based evaluators and additional patterns, see Advanced options for running evals on experiments via code. For the full client.experiments API, see the Python SDK Experiments reference.

View eval results

Results appear in the experiment table alongside each row. Compare multiple experiment runs side by side, filter by eval label, and drill into individual rows to inspect the task output and evaluator explanation. Open View Eval Trace from an experiment row when you want the evaluated trace. For per-example outputs and eval explanations together, use Compare Experiments and the Table tab.

Compare Experiments table tab showing dataset rows, experiment outputs, Evals column with correct and incorrect tags, and an open popover with score, label, and explanation for an evaluator

The CI/CD workflow

Make your change (prompt update, model swap, logic change).
Run your experiment against the golden dataset.
Score the experiment results with your evaluators.
Compare eval scores against the previous experiment run.
If scores regress, fix before merging.

You can integrate this into GitHub Actions or GitLab CI/CD so it runs automatically on every push or pull request. See GitHub Action basics and GitLab CI/CD basics for full guides.

Define evaluators once in Create evaluators, then attach them to experiments.
After you ship, use Run online evals on traces to monitor production with the same criteria where possible.
Follow the Develop tutorials: Defining the dataset, Run experiments with code evals, Run experiments with LLM judge, and Iteration workflow.

How to Use Arize

Quickstart

Instrument

Observe

Evaluate

Develop

Prompts

Machine Learning

Settings

Security

Run offline evals on experiments

Eval-driven development

Run on an experiment

Evaluation Inputs

Evaluation Outputs

Create an LLM Evaluator

Run Evaluation

Create a Code Evaluator

Custom Code Evaluators

Prebuilt Phoenix Code Evaluators

View eval results

The CI/CD workflow

How to Use Arize

Quickstart

Instrument

Observe

Evaluate

Develop

Prompts

Machine Learning

Settings

Security

​Eval-driven development

​Run on an experiment

​Evaluation Inputs

​Evaluation Outputs

​Create an LLM Evaluator

​Run Evaluation

​Create a Code Evaluator

​Custom Code Evaluators

​Prebuilt Phoenix Code Evaluators

​View eval results

​The CI/CD workflow

​Related workflows

Eval-driven development

Run on an experiment

Evaluation Inputs

Evaluation Outputs

Create an LLM Evaluator

Run Evaluation

Create a Code Evaluator

Custom Code Evaluators

Prebuilt Phoenix Code Evaluators

View eval results

The CI/CD workflow

Related workflows