Mistral vs Llama 2: benchmark on your own data
Mistral was recently launched as the "best 7B model to date". This claim is made on the basis of a number of evals performed by the researchers.
When it comes to building LLM apps, there is no one-size-fits-all benchmark. To maximize the quality of your LLM application, consider building your own benchmark to supplement public benchmarks. This guide describes how to compare Mistral-7B-v0.1 vs Llama 7B using the promptfoo CLI.
The end result is a view that compares the performance of Mistral and Llama side-by-side:
View the final example code here.
Requirements
This guide assumes that you have promptfoo installed. It also requires HuggingFace and Replicate access, but in principle you can follow these instructions for any local LLM.
Set up the config
Initialize a new directory mistral-llama-comparison that will contain our prompts and test cases:
npx promptfoo@latest init mistral-llama-comparison
Now let's start editing promptfooconfig.yaml. Create a list of models we'd like to compare:
providers:
- huggingface:text-generation:mistralai/Mistral-7B-Instruct-v0.1
- replicate:replicate/llama70b-v2-chat:e951f18578850b652510200860fc4ea62b3b16fac280f83ff32282f87bbd2e48
The first provider references the model Mistral-7B-Instruct-v0.1 on HuggingFace. The second provider references the hosted Replicate version of chat-tuned Llama v2, which isn't available on the HuggingFace Inference API.
If you prefer to run against a locally hosted versions of these models, this can be done via LocalAI, Ollama, or Llama.cpp (using the quantized Mistral).
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 Mistral and Llama expect different prompting formats.
First, we'll put the Mistral chat prompt in prompts/mistral_prompt.txt using the special <s> and [INST] tokens that the model was fine-tuned on:
<s>[INST]{{message}}[/INST]
Next, we'll put the Llama chat prompt in prompts/llama_prompt.txt:
User: {{message}}
Assistant:
Now, let's go back to promptfooconfig.yaml and add our prompts. We'll name them mistral_prompt and llama_prompt respectively:
prompts:
prompts/mistral_prompt.txt: mistral_prompt
prompts/llama_prompt.txt: llama_prompt
providers:
- huggingface:text-generation:mistralai/Mistral-7B-Instruct-v0.1:
prompts: mistral_prompt
- replicate:meta/llama-2-7b-chat:8e6975e5ed6174911a6ff3d60540dfd4844201974602551e10e9e87ab143d81e:
prompts: llama_prompt
These prompt files are Nunjucks templates, so you can use if statements, for loops, and filters for more complex prompts.
Configure model parameters
Each model has a config field where you can specify additional parameters. Let's add temperature and max_length for each model:
providers:
- huggingface:text-generation:mistralai/Mistral-7B-Instruct-v0.1:
prompts: mistral_prompt
config:
temperature: 0.01
max_new_tokens: 128
- replicate:meta/llama-2-7b-chat:8e6975e5ed6174911a6ff3d60540dfd4844201974602551e10e9e87ab143d81e:
prompts: llama_prompt
config:
temperature: 0.01
max_new_tokens: 128
Mistral supports HuggingFace text generation parameters whereas Replicate's API has its own set of supported parameters.
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_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 HuggingFace and Replicate providers, be sure to set the following environment variables:
HF_API_TOKEN=your_huggingface_api_key
REPLICATE_API_TOKEN=your_replicate_api_token
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?
Check for correctness
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?
Learn more about setting up test assertions here.
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 JSON, YAML, or CSV by specifying an output file:
npx promptfoo@latest eval -o output.csv
Conclusion
On this limited dataset, Mistral scores 87.50% and Llama2 scores 50%. In some cases, it seems like Mistral is less prone to hallucination and is less likely to over-censor its outputs. But these are just a handful of use cases - far from conclusive.
Ultimately, if you are considering these LLMs for a specific use case, you should eval them for exactly that use case. Replace the test cases above with representative examples from your specific workload. This will create a much for specific and useful benchmark.