benchllm 0.3.0

Tool for testing LLMs

🏋️‍♂️ BenchLLM 🏋️‍♀️

🦾 Continuous Integration for LLM powered applications 🦙🦅🤖

BenchLLM is a Python-based open-source library that streamlines the testing of Large Language Models (LLMs) and AI-powered applications. It measures the accuracy of your model, agents, or chains by validating responses on any number of tests via LLMs.

BenchLLM is actively used at V7 for improving our LLM applications and is now Open Sourced under MIT License to share with the wider community

💡 Get help on Discord or Tweet at us

---

Use BenchLLM to:

• Test the responses of your LLM across any number of prompts.
• Continuous integration for chains like Langchain, agents like AutoGPT, or LLM models like Llama or GPT-4.
• Eliminate flaky chains and create confidence in your code.
• Spot inaccurate responses and hallucinations in your application at every version.

---

⚠️ NOTE: BenchLLM is in the early stage of development and will be subject to rapid changes.

For bug reporting, feature requests, or contributions, please open an issue or submit a pull request (PR) on our GitHub page.

🧪 BenchLLM Testing Methodology

BenchLLM implements a distinct two-step methodology for validating your machine learning models:

1. Testing: This stage involves running your code against any number of expected responses and capturing the predictions produced by your model without immediate judgment or comparison.

2. Evaluation: The recorded predictions are compared against the expected output using LLMs to verify factual similarity (or optionally manually). Detailed comparison reports, including pass/fail status and other metrics, are generated.

This methodical separation offers a comprehensive view of your model's performance and allows for better control and refinement of each step.

🚀 Install

To install BenchLLM we use pip

pip install benchllm

💻 Usage

Start by importing the library and use the @benchllm.test decorator to mark the function you'd like to test:

import benchllm

# Your custom model implementation
def run_my_model(input):
    # Your model's logic goes here.
    return some_result

@benchllm.test(suite="/path/to/test/suite") # If the tests are in the same directory, just use @benchllm.test.
def invoke_model(input: str):
    return run_my_model(input)

Next, prepare your tests. These are YAML/JSON files structured as follows:

input: What's 1+1? Be very terse, only numeric output
expected:
  - 2
  - 2.0

In the above example, the input is the query or instruction that your model will process, and expected contains the potential responses that your model should return. It's important to note that input can be a simple str or a more complex nested dictionary; BenchLLM will extract the type of the input argument in the Python code and load the input field from the YAML file accordingly.

By default, BenchLLM uses OpenAI's GPT-3 model for the semantic evaluator. This requires setting the OPENAI_API_KEY environment variable. If you do not want to use this default evaluator, you can specify an alternative one (discussed in further detail below):

export OPENAI_API_KEY='your-api-key'

Replace 'your-api-key' with your actual OpenAI API key.

To initiate testing, use the bench run command:

$ bench run

By default, the bench run command looks for Python files implementing the @test decorator in the current directory. To target a specific file or folder, specify it directly:

$ bench run path/to/my/file.py or/path/to/folder/with/files

The --retry-count parameter allows BenchLLM to run a test multiple times, useful for models that may have variability in their outputs:

$ bench run --retry-count 5

BenchLLM offers multiple evaluation methods to determine if the prediction matches the test case's expected values. You can use the --evaluator parameter to specify the evaluation method:

There are multiple ways to evaluate if the test functions prediction matches the test cases expected values. By default GPT-3 is used to compare the output. You can use --evaluator to use a different method

• semantic, checks semantic similarity using language models like GPT-3, GPT-3.5, or GPT-4 (--model parameter). Please note, for this evaluator, you need to set the OPENAI_API_KEY environment variable.
• embedding, uses cosine distance between embedded vectors. Please note, for this evaluator, you need to set the OPENAI_API_KEY environment variable.
• string-match, checks if the strings are matching (case insensitive)
• interactive, user manually accepts or fails tests in the terminal
• web, uses pywebio fora simple local web interface

The non interactive evaluators also supports --workers N to run in the evaluations in parallel

$ bench run --evaluator string-match --workers 5

To accelerate the evaluation process, BenchLLM uses a cache. If a (prediction, expected) pair has been evaluated in the past and a cache was used, the evaluation output will be saved for future evaluations. There are several types of caches:

• memory, only caches output values during the current run. This is particularly useful when running with --retry-count N
• file, stores the cache at the end of the run as a JSON file in output/cache.json. This is the default behavior.
• none, does not use any cache.

$ bench run examples --cache memory

When working on developing chains or training agent models, there may be instances where these models need to interact with external functions — for instance, querying a weather forecast or executing an SQL query. In such scenarios, BenchLLM facilitates the ability to mock these functions. This helps you make your tests more predictable and enables the discovery of unexpected function calls.

input: I live in London, can I expect rain today?
expected: ["no"]
calls:
  - name: forecast.get_n_day_weather_forecast
    returns: It's sunny in London.
    arguments:
      location: London
      num_days: 1

In the example above, the function get_n_day_weather_forecast in the forecast module is mocked. In other words, every time this function is invoked, the model will receive "It's sunny in London". BenchLLM also provides warnings if the function is invoked with argument values different from get_n_day_weather_forecast(location=London, num_days=1). Please note, the provision of these argument parameters is optional.

🧮 Eval

While bench run runs each test function and then evaluates their output, it can often be beneficial to separate these into two steps. For example, if you want a person to manually do the evaluation or if you want to try multiple evaluation methods on the same function.

$ bench run --no-eval

This will generate json files in output/latest/predictions Then later you can evaluate them with

$ bench eval output/latest/predictions

🔌 API

For more detailed control, BenchLLM provides an API. You are not required to add YML/JSON tests to be able to evaluate your model. You can instead:

• Instantiate Test objects
• Use a Tester object to generate predictions
• Use an Evaluator object to evaluate your model

from benchllm import StringMatchEvaluator, Test, Tester

# Instantiate your Test objects
tests = [
    Test(input="What's 1+1?", expected=["2", "It's 2"]),
    Test(input="First rule of fight club?", expected=["Do not talk about fight club"]),
]

# Use a Tester object to generate predictions using any test functions
tester = Tester(my_test_function)
tester.add_tests(tests)
predictions = tester.run()

# Use an Evaluator object to evaluate your model
evaluator = StringMatchEvaluator()
evaluator.load(predictions)
results = evaluator.run()

print(results)

If you want to incorporate caching and run multiple parallel evaluation jobs, you can modify your evaluator as follows:

from benchllm.cache import FileCache

...

evaluator = FileCache(StringMatchEvaluator(workers=2), Path("path/to/cache.json"))
evaluator.load(predictions)
results = evaluator.run()

In this example, FileCache is used to enable caching, and the workers parameter of StringMatchEvaluator is set to 2 to allow for parallel evaluations. The cache results are saved in a file specified by Path("path/to/cache.json").

☕️ Commands

• bench add: Add a new test to a suite.
• bench tests: List all tests in a suite.
• bench run: Run all or target test suites.
• bench eval: Runs the evaluation of an existing test run.

🙌 Contribute

BenchLLM is developed for Python 3.10, although it may work with other Python versions as well. We recommend using a Python 3.10 environment and pip >= 23. You can use conda or any other environment manager to set up the environment:

$ conda create --name benchllm python=3.10
$ conda activate benchllm
$ pip install -e ".[dev]"

To run all the examples first install the examples extra dependencies

$ pip install -e ".[examples]"

Contribution steps:

1. Fork the repository.
2. Create a new branch for your changes.
3. Make your changes.
4. Test your changes.
5. Submit a pull request.

We adhere to the PEP8 style guide. Please follow this guide when contributing.

If you need any support, feel free to open an issue on our GitHub page.