LLM evaluation framework
This project enables the creation of your own LLM evaluation framework against popular LLM providers (Anthropic, Google, OpenAI) and cloud providers (Google Cloud) and then output to a formatted text table or HTML table. It is lightweight and easy to read and extend by design.
See Which LLM is the Best? Let's Build a Model Evaluation Framework and Find Out. for additional background.
Setup
Install required packages:
pip install -r requirements.txt
Update config.ini before running to enable APIs and services as needed:
[API Keys]
ANTHROPIC_API_KEY = ADD_HERE
GOOGLE_AI_STUDIO_API_KEY = ADD_HERE
OPENAI_API_KEY = ADD_HERE
[Cloud Configs]
GCP_PROJECT = ADD_HERE
GCP_PROJECT_LOCATION = ADD_HERE
Usage
Run test.py or walk through the notebook BSD_Evals.ipynb to get started:
Once your evaluation has completed you will see an evaluation summary:
Total runtime: 65.56242179870605
Models: 8
Evals: 6
Total Evals: 48
Passed Evals: 26
Failed Evals: 14
Other Evals: 8
An evaluation matrix:
And a runtime matrix:
Additional Configurations
See evals/test_evals.json for an example on how to define your own set of evaluations.
Evaluation types include:
- perfect_exact_match: Expected and actual results must match exactly.
- case_insensitive_match: Same as perfect_exact_match except upper and lowercase letters are treated as a match.
- rouge: Measures the similarity of text passages. This is good for summarization use cases.
- sql_generation_results: For prompts that generate SQL, this takes the next step and executes the generated SQL. Then the SQL result is compared to expected output and must match exactly (perfect_exact_match).
Final output can be defined as a formatted text table or HTML table:
evals.display_results("text") # Formatted text table
evals.display_results("html") # HTML table