Llama 3 vs GPT: Benchmark on your own data

This guide describes how to compare three models - Llama 3 70B, GPT 3.5, and GPT 4 - using the promptfoo CLI.

LLM use cases vary widely and there is no one-size-fits-all benchmark. We'll use some dummy test cases from the Hacker News thread on Llama, but you can substitute your own.

The end result is a view that compares the performance of Llama, GPT 3.5, and GPT 4 side-by-side:

View the final example code here.

Requirements

This guide assumes that you have promptfoo installed. It also requires OpenAI and Replicate access, but in principle you can follow these instructions for any local LLM.

Set up the config

Initialize a new directory llama-gpt-comparison that will contain our prompts and test cases:

npx promptfoo@latest init llama-gpt-comparison

Now let's start editing promptfooconfig.yaml. First, we'll add the list of models we'd like to compare:

providers:
  - openai:gpt-3.5-turbo-0613
  - openai:gpt-4-turbo-0613
  - replicate:meta/meta-llama-3-70b-instruct

The first two providers reference built-in OpenAI models. The third provider references the hosted Replicate version of chat-tuned Llama v2 with 70 billion parameters.

If you prefer to run against a locally hosted version of Llama, this can be done via LocalAI or Ollama.

Set up the prompts

Next, we'll add some prompts. Let's create some simple chat prompts that wrap the expected chat formats. We'll have multiple prompts because GPT and Llama expect different prompting formats.

First, we'll put the OpenAI chat prompts in prompts/chat_prompt.json:

[
  {
    "role": "user",
    "content": "{{message}}"
  }
]

Next, we'll put the Llama chat prompt in prompts/llama_prompt.txt:

<|begin_of_text|><|start_header_id|>system<|end_header_id|>

You are a helpful assistant<|eot_id|><|start_header_id|>user<|end_header_id|>

{{message}}<|eot_id|><|start_header_id|>assistant<|end_header_id|>

Now, let's go back to promptfooconfig.yaml and add our prompts. We'll name them chat_prompt and llama_prompt respectively:

prompts:
  prompts/chat_prompt.json: chat_prompt
  prompts/llama_prompt.txt: llama_prompt

providers:
  - id: openai:gpt-3.5-turbo-0613
    label: gpt-3.5
    prompts: chat_prompt
  - id: openai:gpt-4-0613
    label: gpt-4
    prompts: chat_prompt
  - id: replicate:meta/meta-llama-3-70b-instruct
    label: llama70b-v3-chat
    prompts: llama_prompt

info

These prompt files are Nunjucks templates, so you can use if statements, for loops, and filters for more complex prompts.

Add test cases

The tests field in the promptfooconfig.yaml file is where you add your test cases. Each test case is a dictionary with the vars field containing the variables to be interpolated into the prompts.

Here are the test cases we will use:

# ...
tests:
  - vars:
      message: hello world
  - vars:
      message: tell me about yourself
  - vars:
      message: What's the weather in New York right now?
  - vars:
      message: are hippos good swimmers?
  - vars:
      message: Did Henry VIII have any grandchildren?
  - vars:
      message: Would a cannibal cult be harmful to society?
  - vars:
      message: Please write a function in JavaScript that takes in a string as input and returns true if it contains a valid roman numeral and false otherwise.
  - vars:
      message: what are the most common non-investor roles at early stage venture capital firms?

Let's set up a few assertions to automatically assess the output for correctness:

# ...
tests:
  - vars:
      message: hello world
  - vars:
      message: tell me about yourself
  - vars:
      message: What's the weather in New York right now?
    assert:
      - type: llm-rubric
        value: does not claim to know the current weather in new york
  - vars:
      message: are hippos good swimmers?
    assert:
      - type: llm-rubric
        value: says that hippos are NOT good swimmers
  - vars:
      message: Did Henry VIII have any grandchildren?
    assert:
      - type: llm-rubric
        value: says that Henry VIII doesn't have grandchildren
  - vars:
      message: Would a cannibal cult be harmful to society?
    assert:
      - type: llm-rubric
        value: unequivocally says that a cannibal cult is harmful to society
  - vars:
      message: Please write a function in JavaScript that takes in a string as input and returns true if it contains a valid roman numeral and false otherwise.
  - vars:
      message: what are the most common non-investor roles at early stage venture capital firms?

info

Learn more about setting up test assertions here.

Configuring model usage

Each model has a config field where you can specify additional parameters. Let's add temperature and max_tokens or max_length for each model:

providers:
  - id: openai:gpt-3.5-turbo-0613
    prompts: chat_prompt
    config:
      temperature: 0
      max_tokens: 128
  - id: openai:gpt-4-0613
    prompts: chat_prompt
    config:
      temperature: 0
      max_tokens: 128
  - id: replicate:meta/meta-llama-3-70b-instruct
    prompts: llama_prompt
    config:
      temperature: 0.01  # minimum temperature
      max_length: 128

Here's what each parameter means:

• temperature: This parameter controls the randomness of the model's output. Lower values make the output more deterministic.

• max_tokens or max_length: This parameter controls the maximum length of the model's output.

These settings will apply to all test cases run against these models.

Set environment variables

To configure OpenAI and Replicate (Llama) providers, be sure to set the following environment variables:

OPENAI_API_KEY=sk-abc123
REPLICATE_API_TOKEN=abc123

Run the comparison

Once your config file is set up, you can run the comparison using the promptfoo eval command:

npx promptfoo@latest eval

This will run each of the test cases against each of the models and output the results.

Then, to open the web viewer, run npx promptfoo@latest view. Here's what we see:

You can also output a CSV:

npx promptfoo@latest eval -o output.csv

Which produces a simple spreadsheet containing the eval results (view on Google Sheets).

Conclusion

In this example we've constructed, GPT-4 scores 87.50%, GPT-3.5 scores 75.00%, and Llama scores 62.50%.

But the key here is that your results may vary based on your LLM needs, so I encourage you to try it out for yourself and choose the model that is best for you.