yu-jeffy / GreedLlama

2 stars 0 forks source link

GreedLlama

A Study of Profit-Tuned LLMs in Morally Ambiguous Decision-Making

Now published on arxiv

arxiv 2404.02934

Brief Description:

GreedLlama is a study focused on evaluating the behavioral outcomes of fine-tuned language models with a specific inclination toward profit maximization. This study targets the performance of a modified version of a language model—referred to as "GreedLlama"—against a standard model. We're currently conducting the first phase of analysis, examining how these distinct configurations interact with moral reasoning scenarios.

We employ sentiment analysis techniques to assess the differences in the answers provided by both models when evaluated against a moral reasoning dataset.

Objective:

The objective of GreedLlama is to shed light on the implications of deploying profit-focused language models within business environments. Our study addresses the potential ethical trade-offs and the necessity for multi-layered oversight when integrating these models into decision-making processes that can significantly impact human lives and well-being.

Current Status:

The study is completed.

Contact:

For more information or to contribute, please contact the project maintainers.