Summary of “all That Glitters”: Approaches to Evaluations with Unreliable Model and Human Annotations, by Michael Hardy

“All that Glitters”: Approaches to Evaluations with Unreliable Model and Human Annotations

by Michael Hardy

First submitted to arxiv on: 23 Nov 2024

GrooveSquid.com Paper Summaries

GrooveSquid.com’s goal is to make artificial intelligence research accessible by summarizing AI papers in simpler terms. Each summary below covers the same AI paper, written at different levels of difficulty. The medium difficulty and low difficulty versions are original summaries written by GrooveSquid.com, while the high difficulty version is the paper’s original abstract. Feel free to learn from the version that suits you best!

Summary difficulty	Written by	Summary
High	Paper authors	High Difficulty Summary Read the original abstract here
Medium	GrooveSquid.com (original content)	Medium Difficulty Summary The paper presents methods for evaluating label quality and addressing issues with “gold” and “ground truth” human-mediated labels in model evaluation. It explores two Large Language Model (LLM) architecture families – encoders and GPT decoders – to automate the task of annotating classroom teaching quality, an important but expensive task currently only done by humans. The study demonstrates that using standard metrics can mask label and model quality issues, revealing spurious correlations and non-random racial biases across models and humans. Novel evaluation methods are proposed to assess label quality across six dimensions: Concordance, Confidence, Validity, Bias, Fairness, and Helpfulness.
Low	GrooveSquid.com (original content)	Low Difficulty Summary The paper tries to solve a problem with how we measure the quality of labels that humans use to teach models new things. Right now, it’s hard to know if the labels are accurate or not because they can be wrong sometimes. The study uses special kinds of computer models called Large Language Models (LLMs) to see if they can do better than humans at labeling classroom teaching. It found that some LLMs did really well, but others had problems with bias and accuracy. The paper also shows how using these LLMs in a way that involves human feedback could make things worse or better depending on the model.