Code for "Leveraging Large Language Models for Predictive Analysis of Human Misery"

Dataset - https://huggingface.co/datasets/bansalaman18/misery-index

• misery_index_game_show.py - code for running the LLM as a game show contestant

• utils.py - functions for generating responsed from LLM functions

• eval.py - evaluating the responses

• Misery_Data.csv - data containing miserable situations, their annotated misery index values, contestant responses, and other metadata

  • Run python misery_index_game_show.py <SEED> <MODEL_NAME>

    • <SEED> - 12, 123, 1234
    • <MODEL_NAME> - gpt-3.5-turbo, gpt-4, gpt-4-turbo, gpt-4o, gpt-4o-mini, o1, o1-mini, gemini-1.5-pro

Progress

• Game-show setting

Column headers are seeds.

	12	123	1234
gpt-3.5-turbo	✅	✅	✅
gpt-4	✅	❌	❌
gpt-4-turbo	✅	✅	✅
gpt-4o-mini	✅	✅	✅
gpt-4o	✅	✅	✅
o1-preview	❌	❌	❌
o1-mini	❌	❌	❌
gemini-1.5-pro	❌	❌	❌

• Game-show setting with chain-of-thought

• Directly predicting the misery index

• Binary Comparisons between situations

Links used for gathering data

• https://bobbymgsk.wordpress.com/category/the-misery-index/
• https://jericho.blog/2021/02/03/the-misery-index-data/ - contains most of the data in a spreadsheet - https://docs.google.com/spreadsheets/d/151WjFwDdhIURf48subj6SDOdra0XVIEo0xulnBMMfRo/edit#gid=1169151367