Author(s): Meng Li

Originally published on Towards AI.

How to Make AI Understand Your Inner Voice?
Created by Meng Li

Recently, I have been researching large language models with the goal of using them to score samples.

Initially, I attempted to have the model rate on a scale from 1 to 5, but it consistently gave out only the lowest or highest scores, lacking nuance in its evaluations.

So, I switched tactics, moving away from numerical scores to descriptive phrases ranging from “very poor” to “excellent.”

This change immediately diversified the model’s ratings, moving away from simplistic extreme judgments.

I went a step further and tried scoring with a percentage system, from 1 to 100. This broader range allowed the model to more precisely assess the quality of samples.

These experiments were essentially about reward models in reinforcement learning.

Through these experiments, I discovered that different scoring methods significantly impact the model’s performance.

Now, I want to share a case study with everyone.

We will explore how these various scoring methods differ in practical applications.

The Likert Scale features five levels, ranging from 1 to 5, representing attitudes from “strongly disagree” to “strongly agree.”

I have created a template of prompt words based on this scale.

Evaluate the score of a novel excerpt; if the score is low, modify it based on keywords.

Please assess the overall quality of… Read the full blog for free on Medium.

Join thousands of data leaders on the AI newsletter. Join over 80,000 subscribers and keep up to date with the latest developments in AI. From research to projects and ideas. If you are building an AI startup, an AI-related product, or a service, we invite you to consider becoming a sponsor.

Published via Towards AI