Build an evaluation

Learn how to build an evaluation pipeline with Weave Models and Evaluations

This tutorial walks you through building an end-to-end evaluation pipeline in Weave so you can measure and track the quality of an LLM application as you iterate on it. Evaluations help you compare changes against a consistent set of examples and detect regressions before they reach users. This tutorial targets developers who build LLM-powered applications and want a repeatable way to test them.

Weave provides built-in support for tracking evaluations with Model and Evaluation classes. The APIs make minimal assumptions, so they fit a range of use cases.

What you’ll learn

This guide shows you how to:

• Set up a Model.
• Create a dataset to test an LLM’s responses against.
• Define a scoring function to compare model output to expected outputs.
• Run an evaluation that tests the model against the dataset using the scoring function and an additional built-in scorer.
• View the results of the evaluation in the Weave UI.

By the end, you’ll have a working evaluation pipeline that scores an example model against a dataset and logs the results to Weave.

Prerequisites

• A W&B account
• Python 3.10+ or Node.js 18+
• Required packages installed:
  • Python: pip install weave openai
  • TypeScript: npm install weave openai
• An OpenAI API key set as an environment variable.

Import the necessary libraries and functions

Import the following libraries into your script:

Python

import json
import openai
import asyncio
import weave
from weave.scorers import MultiTaskBinaryClassificationF1

import * as weave from 'weave';
import OpenAI from 'openai';

Build a Model

With the libraries in place, the next step is to define the model you want to evaluate.

In Weave, Models are objects that capture both the behavior of your model or agent (logic, prompt, parameters) and its versioned metadata (parameters, code, micro-config) so you can track, compare, evaluate, and iterate reliably.

When you instantiate a Model, Weave automatically captures its configuration and behaviors and updates the version when changes occur. This lets you track its performance over time as you iterate on it.

To declare a Model, subclass Model and implement a predict function definition that takes one example and returns the response.

The following example model uses OpenAI to extract the names, colors, and flavors of alien fruits from input sentences.

Python

class ExtractFruitsModel(weave.Model):
    model_name: str
    prompt_template: str

    @weave.op()
    async def predict(self, sentence: str) -> dict:
        client = openai.AsyncClient()

        response = await client.chat.completions.create(
            model=self.model_name,
            messages=[
                {"role": "user", "content": self.prompt_template.format(sentence=sentence)}
            ],
        )
        result = response.choices[0].message.content
        if result is None:
            raise ValueError("No response from model")
        parsed = json.loads(result)
        return parsed

// Note: weave.Model is not supported in TypeScript yet.
// Instead, wrap your model-like function with weave.op

import * as weave from 'weave';
import OpenAI from 'openai';

const openaiClient = new OpenAI();

const model = weave.op(async function myModel({datasetRow}) {
  const prompt = `Extract fields ("fruit": <str>, "color": <str>, "flavor") from the following text, as json: ${datasetRow.sentence}`;
  const response = await openaiClient.chat.completions.create({
    model: 'gpt-3.5-turbo',
    messages: [{ role: 'user', content: prompt }],
    response_format: { type: 'json_object' }
  });
  return JSON.parse(response.choices[0].message.content);
});

The ExtractFruitsModel class inherits from (or subclasses) weave.Model so Weave can track the instantiated object. @weave.op decorates the predict function to track its inputs and outputs.

You can instantiate Model objects like this:

Python

# Set your team and project name
weave.init('[YOUR-TEAM]/eval_pipeline_quickstart')

model = ExtractFruitsModel(
    model_name='gpt-3.5-turbo-1106',
    prompt_template='Extract fields ("fruit": <str>, "color": <str>, "flavor": <str>) from the following text, as json: {sentence}'
)

sentence = "There are many fruits that were found on the recently discovered planet Goocrux. There are neoskizzles that grow there, which are purple and taste like candy."

print(asyncio.run(model.predict(sentence)))
# if you're in a Jupyter Notebook, run:
# await model.predict(sentence)

await weave.init('eval_pipeline_quickstart');

const sentence = "There are many fruits that were found on the recently discovered planet Goocrux. There are neoskizzles that grow there, which are purple and taste like candy.";

const result = await model({ datasetRow: { sentence } });

console.log(result);

Create a dataset

With a Model defined, you now need a dataset to evaluate it against. A Dataset is a collection of examples stored as a Weave object. Publishing the dataset to Weave versions it and makes it reusable across evaluation runs.

The following example dataset defines three example input sentences and their correct answers (labels), and then formats them in a JSON table format that scoring functions can read.

This example builds a list of examples in code, but you can also log them one at a time from your running application.

Python

sentences = ["There are many fruits that were found on the recently discovered planet Goocrux. There are neoskizzles that grow there, which are purple and taste like candy.",
"Pounits are a bright green color and are more savory than sweet.",
"Finally, there are fruits called glowls, which have a very sour and bitter taste which is acidic and caustic, and a pale orange tinge to them."]
labels = [
    {'fruit': 'neoskizzles', 'color': 'purple', 'flavor': 'candy'},
    {'fruit': 'pounits', 'color': 'bright green', 'flavor': 'savory'},
    {'fruit': 'glowls', 'color': 'pale orange', 'flavor': 'sour and bitter'}
]
examples = [
    {'id': '0', 'sentence': sentences[0], 'target': labels[0]},
    {'id': '1', 'sentence': sentences[1], 'target': labels[1]},
    {'id': '2', 'sentence': sentences[2], 'target': labels[2]}
]

const sentences = [
  "There are many fruits that were found on the recently discovered planet Goocrux. There are neoskizzles that grow there, which are purple and taste like candy.",
  "Pounits are a bright green color and are more savory than sweet.",
  "Finally, there are fruits called glowls, which have a very sour and bitter taste which is acidic and caustic, and a pale orange tinge to them."
];
const labels = [
  { fruit: 'neoskizzles', color: 'purple', flavor: 'candy' },
  { fruit: 'pounits', color: 'bright green', flavor: 'savory' },
  { fruit: 'glowls', color: 'pale orange', flavor: 'sour and bitter' }
];
const examples = sentences.map((sentence, i) => ({
  id: i.toString(),
  sentence,
  target: labels[i]
}));

Then create your dataset using the weave.Dataset() class and publish it:

Python

weave.init('eval_pipeline_quickstart')
dataset = weave.Dataset(name='fruits', rows=examples)
weave.publish(dataset)

import * as weave from 'weave';
await weave.init('eval_pipeline_quickstart');
const dataset = new weave.Dataset({
  name: 'fruits',
  rows: examples
});
await dataset.save();

Define custom scoring functions

Now that you have a model and a dataset, you need a way to measure how well the model performs on each example. Scoring functions compare the model’s output to the expected target and produce the metrics that an evaluation reports.

When you use Weave evaluations, Weave expects a target to compare output against. The following scoring function takes two dictionaries (target and output) and returns a dictionary of boolean values that indicate whether the output matches the target. The @weave.op() decorator enables Weave to track the scoring function’s execution.

Python

@weave.op()
def fruit_name_score(target: dict, output: dict) -> dict:
    return {'correct': target['fruit'] == output['fruit']}

import * as weave from 'weave';

const fruitNameScorer = weave.op(
  function fruitNameScore({target, output}) {
    return { correct: target.fruit === output.fruit };
  }
);

For more information on making your own scoring function, see the Scorers guide.

In some applications, you may want to create custom Scorer classes. For example, you might create a standardized LLMJudge class with specific parameters (such as chat model or prompt), specific row scoring, and aggregate score calculation. For more information, see the tutorial on defining a Scorer class in Model-based evaluation of RAG applications.

Use a built-in scorer and run the evaluation

With the model, dataset, and a custom scorer in place, you’re ready to wire everything together into an evaluation run.

Along with custom scoring functions, you can also use Weave’s built-in scorers. In the following evaluation, weave.Evaluation() uses the fruit_name_score function defined in the previous section and the built-in MultiTaskBinaryClassificationF1 scorer, which computes F1 scores.

The following example runs an evaluation of ExtractFruitsModel on the fruits dataset using the two scoring functions and logs the results to Weave.

Python

weave.init('eval_pipeline_quickstart')

evaluation = weave.Evaluation(
    name='fruit_eval',
    dataset=dataset, 
    scorers=[
        MultiTaskBinaryClassificationF1(class_names=["fruit", "color", "flavor"]), 
        fruit_name_score
    ],
)
print(asyncio.run(evaluation.evaluate(model)))
# if you're in a Jupyter Notebook, run:
# await evaluation.evaluate(model)

import * as weave from 'weave';

await weave.init('eval_pipeline_quickstart');

const evaluation = new weave.Evaluation({
  name: 'fruit_eval',
  dataset: dataset,
  scorers: [fruitNameScorer],
});
const results = await evaluation.evaluate(model);
console.log(results);

If you’re running from a Python script, you’ll need to use asyncio.run. However, if you’re running from a Jupyter notebook, you can use await directly.

Complete example

Complete evaluation pipeline in one script:

Python

import json
import asyncio
import openai
import weave
from weave.scorers import MultiTaskBinaryClassificationF1

# Initialize Weave once
weave.init('eval_pipeline_quickstart')

# 1. Define Model
class ExtractFruitsModel(weave.Model):
    model_name: str
    prompt_template: str

    @weave.op()
    async def predict(self, sentence: str) -> dict:
        client = openai.AsyncClient()
        response = await client.chat.completions.create(
            model=self.model_name,
            messages=[{"role": "user", "content": self.prompt_template.format(sentence=sentence)}],
        )
        result = response.choices[0].message.content
        if result is None:
            raise ValueError("No response from model")
        return json.loads(result)

# 2. Instantiate model
model = ExtractFruitsModel(
    model_name='gpt-3.5-turbo-1106',
    prompt_template='Extract fields ("fruit": <str>, "color": <str>, "flavor": <str>) from the following text, as json: {sentence}'
)

# 3. Create dataset
sentences = ["There are many fruits that were found on the recently discovered planet Goocrux. There are neoskizzles that grow there, which are purple and taste like candy.",
"Pounits are a bright green color and are more savory than sweet.",
"Finally, there are fruits called glowls, which have a very sour and bitter taste which is acidic and caustic, and a pale orange tinge to them."]
labels = [
    {'fruit': 'neoskizzles', 'color': 'purple', 'flavor': 'candy'},
    {'fruit': 'pounits', 'color': 'bright green', 'flavor': 'savory'},
    {'fruit': 'glowls', 'color': 'pale orange', 'flavor': 'sour and bitter'}
]
examples = [
    {'id': '0', 'sentence': sentences[0], 'target': labels[0]},
    {'id': '1', 'sentence': sentences[1], 'target': labels[1]},
    {'id': '2', 'sentence': sentences[2], 'target': labels[2]}
]

dataset = weave.Dataset(name='fruits', rows=examples)
weave.publish(dataset)

# 4. Define scoring function
@weave.op()
def fruit_name_score(target: dict, output: dict) -> dict:
    return {'correct': target['fruit'] == output['fruit']}

# 5. Run evaluation
evaluation = weave.Evaluation(
    name='fruit_eval',
    dataset=dataset,
    scorers=[
        MultiTaskBinaryClassificationF1(class_names=["fruit", "color", "flavor"]),
        fruit_name_score
    ],
)
print(asyncio.run(evaluation.evaluate(model)))

import * as weave from 'weave';
import OpenAI from 'openai';

// Initialize Weave once
await weave.init('eval_pipeline_quickstart');

// 1. Define Model
// Note: weave.Model is not supported in TypeScript yet.
// Instead, wrap your model-like function with weave.op
const openaiClient = new OpenAI();

const model = weave.op(async function myModel({datasetRow}) {
  const prompt = `Extract fields ("fruit": <str>, "color": <str>, "flavor": <str>) from the following text, as json: ${datasetRow.sentence}`;
  const response = await openaiClient.chat.completions.create({
    model: 'gpt-3.5-turbo',
    messages: [{ role: 'user', content: prompt }],
    response_format: { type: 'json_object' }
  });
  return JSON.parse(response.choices[0].message.content);
});

// 2. Create dataset
const sentences = [
  "There are many fruits that were found on the recently discovered planet Goocrux. There are neoskizzles that grow there, which are purple and taste like candy.",
  "Pounits are a bright green color and are more savory than sweet.",
  "Finally, there are fruits called glowls, which have a very sour and bitter taste which is acidic and caustic, and a pale orange tinge to them."
];
const labels = [
  { fruit: 'neoskizzles', color: 'purple', flavor: 'candy' },
  { fruit: 'pounits', color: 'bright green', flavor: 'savory' },
  { fruit: 'glowls', color: 'pale orange', flavor: 'sour and bitter' }
];
const examples = sentences.map((sentence, i) => ({
  id: i.toString(),
  sentence,
  target: labels[i]
}));

const dataset = new weave.Dataset({
  name: 'fruits',
  rows: examples
});
await dataset.save();

// 3. Define scoring function
const fruitNameScorer = weave.op(
  function fruitNameScore({target, output}) {
    return { correct: target.fruit === output.fruit };
  }
);

// 4. Run evaluation
const evaluation = new weave.Evaluation({
  name: 'fruit_eval',
  dataset: dataset,
  scorers: [fruitNameScorer],
});
const results = await evaluation.evaluate(model);
console.log(results);

View your evaluation results

After the evaluation completes, you can inspect each prediction and scorer result in the Weave UI. Weave automatically captures traces of each prediction and score. To view the results, click the link that the evaluation prints.

Learn more about Weave evaluations

You now have a complete evaluation pipeline. To go deeper into Weave’s evaluation features, see the following resources:

• Learn more about how to build and use scorers.
• Check out Weave’s built-in scoring functions.
• Learn about Model-Based Evaluation for using LLMs as judges.

Next steps

Build a RAG application to learn about evaluating retrieval-augmented generation.