langcheck 0.8.0

Simple, Pythonic building blocks to evaluate LLM-based applications

Simple, Pythonic building blocks to evaluate LLM applications.

Install • Examples • Quickstart • Docs • 日本語 • 中文 • Deutsch

Install

# Install English metrics only
pip install langcheck

# Install English and Japanese metrics
pip install langcheck[ja]

# Install metrics for all languages (requires pip 21.2+)
pip install --upgrade pip
pip install langcheck[all]

Having installation issues? See the FAQ.

Examples

Evaluate Text

Use LangCheck's suite of metrics to evaluate LLM-generated text.

import langcheck

# Generate text with any LLM library
generated_outputs = [
    'Black cat the',
    'The black cat is sitting',
    'The big black cat is sitting on the fence'
]

# Check text quality and get results as a DataFrame (threshold is optional)
langcheck.metrics.fluency(generated_outputs) > 0.5

It's easy to turn LangCheck metrics into unit tests, just use assert:

assert langcheck.metrics.fluency(generated_outputs) > 0.5

LangCheck includes several types of metrics to evaluate LLM applications. Some examples:

Type of Metric	Examples	Languages
Reference-Free Text Quality Metrics	toxicity(generated_outputs) sentiment(generated_outputs) ai_disclaimer_similarity(generated_outputs)	EN, JA, ZH, DE
Reference-Based Text Quality Metrics	semantic_similarity(generated_outputs, reference_outputs) rouge2(generated_outputs, reference_outputs)	EN, JA, ZH, DE
Source-Based Text Quality Metrics	factual_consistency(generated_outputs, sources)	EN, JA, ZH, DE
Query-Based Text Quality Metrics	answer_relevance(generated_outputs, prompts)	EN, JA
Text Structure Metrics	is_float(generated_outputs, min=0, max=None) is_json_object(generated_outputs)	All Languages
Pairwise Text Quality Metrics	pairwise_comparison(generated_outputs_a, generated_outputs_b, prompts)	EN, JA

Visualize Metrics

LangCheck comes with built-in, interactive visualizations of metrics.

# Choose some metrics
fluency_values = langcheck.metrics.fluency(generated_outputs)
sentiment_values = langcheck.metrics.sentiment(generated_outputs)

# Interactive scatter plot of one metric
fluency_values.scatter()

# Interactive scatter plot of two metrics
langcheck.plot.scatter(fluency_values, sentiment_values)

# Interactive histogram of a single metric
fluency_values.histogram()

Augment Data

Text augmentations can automatically generate reworded prompts, typos, gender changes, and more to evaluate model robustness.

For example, to measure how the model responds to different genders:

male_prompts = langcheck.augment.gender(prompts, to_gender='male')
female_prompts = langcheck.augment.gender(prompts, to_gender='female')

male_generated_outputs = [my_llm_app(prompt) for prompt in male_prompts]
female_generated_outputs = [my_llm_app(prompt) for prompt in female_prompts]

langcheck.metrics.sentiment(male_generated_outputs)
langcheck.metrics.sentiment(female_generated_outputs)

Unit Testing

You can write test cases for your LLM application using LangCheck metrics.

For example, if you only have a list of prompts to test against:

from langcheck.utils import load_json

# Run the LLM application once to generate text
prompts = load_json('test_prompts.json')
generated_outputs = [my_llm_app(prompt) for prompt in prompts]

# Unit tests
def test_toxicity(generated_outputs):
    assert langcheck.metrics.toxicity(generated_outputs) < 0.1

def test_fluency(generated_outputs):
    assert langcheck.metrics.fluency(generated_outputs) > 0.9

def test_json_structure(generated_outputs):
    assert langcheck.metrics.validation_fn(
        generated_outputs, lambda x: 'myKey' in json.loads(x)).all()

Monitoring

You can monitor the quality of your LLM outputs in production with LangCheck metrics.

Just save the outputs and pass them into LangCheck.

production_outputs = load_json('llm_logs_2023_10_02.json')['outputs']

# Evaluate and display toxic outputs in production logs
langcheck.metrics.toxicity(production_outputs) > 0.75

# Or if your app outputs structured text
langcheck.metrics.is_json_array(production_outputs)

Guardrails

You can provide guardrails on LLM outputs with LangCheck metrics.

Just filter candidate outputs through LangCheck.

# Get a candidate output from the LLM app
raw_output = my_llm_app(random_user_prompt)

# Filter the output before it reaches the user
while langcheck.metrics.contains_any_strings(raw_output, blacklist_words).any():
    raw_output = my_llm_app(random_user_prompt)