evaluation/README.md

Evaluation

This module contains utilities and scripts for empirical, numerical evaluation of model outputs. Much of the code written in this module is guided by methodologies described in this paper (Watanabe, 2023).

Data formatting

Using the evaluate_model.py script, evaluation can be run on an existing json database of outputs. Database must contain a list of song containers in the following format:

[
    { // container for song 1
        "id": (int) ...,
        "prompt": (str) ...,                  //optional
        "model_response": (str) ...,
        "target_response": (str) ...          //optional
    },
    { // container for song 2
        "id": (int) ..., // unique from song 1
        "prompt": (str) ...,                  //optional
        "model_response": (str) ...,
        "target_response": (str) ...          //optional
    },
    .
    .
    .
]

Quick start

Run the following script to quick start an evaluation:

py evaluation/evaluate_model.py <path_to_your_database> 

To run different measures, consider passing in a list of measures into the --measures argument:

py evaluation/evaluate_model.py <path_to_your_database> --measures diversity meter syllable