Diversity Measures: Domain Independent Proxies for Failure in Language Model Queries

Computing Infrastructure used

Hardware:

• CPU: 9th Gen Intel(R) Core(TM) i7-9750H, 2.60GHz, 6 Cores, 12 Logical Processors
• GPU: Intel(R) UHD Graphics 630
• Memory: 12GB RAM
• Operating System: Windows 10

Software:

• Python Version: 3.11.4

Relevant Libraries

• Scikit-learn: 1.2.2
• Tensorflow: 2.12.0
• Sentence Transformers: 2.12.0
• Imbalanced Learn: 0.11.0
• Pandas: 2.0.1
• OpenAI: 0.27.2

Instructions for running experiment

1. Environment setup

Install the required packages from PyPI:

pip install -r requirements.txt

2. Data collection

In this step, GPT-3.5's responses are collected and evaluated.
First, create an API key on OpenAI.
Then, create a .env file with the following contents:

OPENAI_ORGANIZATION=[YOUR OPENAI ORGANIZATION ID]
OPENAI_API_KEY=[YOUR OPENAI API KEY]

Then, run the following script in Powershell:

scripts/powershell/data-collection

Results from data collection and analysis are stored in ./cache. We have provided the prefilled since data collection might be quite expensive. In order to run a fresh test, the files in ./cache and its subfolders must first be removed.
To do that, run the following script in Powershell:

scripts/powershell/clear-cache

3. Data Analysis

Data analysis is performed in various Jupyter Notebooks.

(Section 5.2) Shannon Entropy diversity measures are evaluated on DRAW-1K, CSQA and Last Letters dataset with 5 different temperature settings in the following notebook:

scripts/analysis/eval_entropy.ipynb

(Section 5.2) Gini impurity diversity measures are evaluated on DRAW-1K, CSQA and Last Letters dataset with 5 different temperature settings in the following notebook:

scripts/analysis/eval_gini.ipynb

(Section 5.3) Centroid-based diversity measures are evaluated on DRAW-1K, CSQA and Last Letters dataset with 5 different temperature settings in the following notebook:

scripts/analysis/eval_centroid.ipynb

(Section 5.4) Experiments regarding few-shot prompting are performed in the following notebook:

scripts/analysis/few_shot.ipynb

(Section 5.5) Experiments regarding few-shot chain-of-thought prompting are performed in the following notebook. This experiment is only performed on DRAW-1K since it was the only dataset which provided intermediary steps which allows for chain-of-thought style few-shot prompting:

scripts/analysis/few_shot_cot.ipynb

(Section 5.6) The effects of ablating various diversity measures for the 10 Layer Multi-Perceptron model are analyzed in the following notebook:

scripts/analysis/ablation_test.ipynb

(Section 5.6) The performance of various machine learning models are tried in the following notebook:

scripts/analysis/classifier_analysis.ipynb

(Section 5.6) The precision-recall curves for the 10 Layer Multi-Perceptron model are produced in the following notebook:

scripts/analysis/pr_curves.ipynb