Skip to content

Better instruction following for large language models

License

Notifications You must be signed in to change notification settings

compatibl/convince

Repository files navigation

CompatibL Convince Package (Community Edition)

GitHub PyPI - Downloads PyPI - Python Version PyPI PRs Welcome

Summary

Better instruction following for large language models

Experiment results and completions can be examined at:

tests/cl/convince/experiments

To run the experiments, follow these steps:

Setting up LLM keys and models

  1. Copy the file called sample_secrets.yaml in project root to .secrets.yaml (note the dot in front) and replace placeholders for each LLM key with your own key, which can be obtained from the LLM providers. Alternatively, they can be set up using .env or envvars (running without the key will print instructions)
  2. Edit the list of models in stub_llms.py if you prefer to run with a different set

Running the experiments

  1. IMPORTANT: Delete files *.completions.* and *.expected.* stored in subdirectories of the test location before running the tests
  2. Run tests in tests/cl/convince/experiments under pytest to generate new results.

Examining the results

  1. The tests will generate new completions and expected test results using your LLM keys and you will be able to examine them using git diff.
  2. Both completions and expected results will change compared to the saved results due to inherent variability of LLM response, however their statistics should remain the same

Copyright

Each individual contributor holds copyright over their contributions to the project. The project versioning is the sole means of recording all such contributions and copyright details. Specifying corporate affiliation or work email along with the commit shall have no bearing on copyright ownership and does not constitute copyright assignment to the employer. Submitting a contribution to this project constitutes your acceptance of these terms.

Because individual contributions are often changes to the existing code, copyright notices in project files must specify The Project Contributors and never an individual copyright holder.

About

Better instruction following for large language models

Resources

License

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published

Contributors 3

  •  
  •  
  •  

Languages