Overview

This script is designed to generate explanations using a language model based on specified parameters and datasets. It allows users to customize various aspects of the explanation generation process through command-line arguments.

Requirements

• Python 3.12
• Required Python packages (install using pip install -r requirements.txt if a requirements file is provided)

Usage

Run the script from the command line:

python main.py [options]

Command-Line Arguments

The script accepts the following command-line arguments:

--parameters

• Type: str
• Default: '0.5'
• Description: Parameter value used in the explanation generation process.

--temperature

• Type: float
• Default: 0.1
• Description: Controls the randomness of the language model's output. A lower value makes the output more deterministic, while a higher value increases randomness.

--top_p

• Type: float
• Default: 0.8
• Description: Implements nucleus sampling by selecting tokens with a cumulative probability up to top_p. This controls the diversity of the output.

--dataset

• Type: str
• Default: 'cora'
• Description: The name of the dataset to be used. Ensure the dataset is available in your environment.

--max_tokens

• Type: int
• Default: 2048
• Description: The maximum number of tokens to generate in the output.

--repetition_penalty

• Type: float
• Default: 1.05
• Description: Penalty applied to reduce the likelihood of repeating the same token. Values greater than 1.0 discourage repetition.

--explainer

• Type: str
• Default: 'cf-gnnfeatures'
• Description: Specifies the explanation method to be used. Options depend on the implementations available in the src.llms module.

Examples

Running with Default Parameters

python main.py

Running with Custom Parameters

python main.py --parameters '0.7' --temperature 0.5 --top_p 0.9 --dataset 'pubmed' --max_tokens 1024 --repetition_penalty 1.1 --explainer 'your_explainer'

You can run the experiments using the experiments script.

bash experiments.sh

To evaluate the results type:

python src/evaluation.py

Data Folder

The folder data contains data coming from CF-GNNExplainer and CF-GNNFeatures explainers using the graph incident format. The explainers are NOT included in this repo!