arXiv-2022 Holistic Evaluation of Language Models

2022-11-09 · arXiv· NLP / Agents · LLM, NLP

Paper: Holistic Evaluation of Language Models

HELM: Holistic Evaluation of Language Models

Summary

InstructGPT performs best overall.
Open-source models still lag behind closed-source models to a noticeable degree.
Larger models are generally better; only above roughly 50B parameters can a model perform well in a given domain.
Prompting helps current models, but performance is highly sensitive to prompt design.

Abstract

Language models have become the foundation of modern NLP systems, yet evaluation of their capabilities, limitations, and risks remains insufficient. The work first taxonomizes potential application scenarios and evaluation criteria; evaluates seven dimensions (accuracy, calibration, robustness, fairness, bias, toxicity, and efficiency); benchmarks 30 language models across 42 application scenarios; and covers 96% of the scenario space.

Introduction

Core Scenarios

For each setting: what task, what domain the data comes from, who constructed it, when it was collected, and so on.

Taxonomy

Many tasks are not fully labeled even in academic benchmarks; there are many relatively niche tasks.

Languages are mainly English and Chinese.

Question answering

One form is open-ended QA; another is closed-set QA (choose from a given list of answers).

See the paper for specific datasets.

Information retrieval

Given a query q and a corpus C, return top-k documents—essentially a ranking problem.

Ranking uses a point-wise approach: concatenate passage c with query q and score the probability that the passage contains an answer to the query; the model’s probability of outputting “yes” is used as the score.

Summarization

Sequence-to-sequence generation.

The focus is abstractive ability rather than extractive copying, which also makes evaluation harder.

Abstraction and extraction are partly in tension: less extraction can hurt factual faithfulness; the goal is to keep generated content as correct as possible.

Sentiment analysis

User reviews of products.

Toxicity detection

Some utterances are acceptable in one cultural context but not in another; the task is to judge whether text is toxic.

Miscellaneous text classification

Various heterogeneous text classification tasks.

General Metrics

Metrics should not be tied to a single scenario, so the framework mainly uses perturbation-based evaluation.

Accuracy

General: exact match, quasi-exact match, F1

Information Retrieval: RR@K, NDCG@K

summarization: ROUGE-2

Language: BPB

Reasoning: F1, exact match

Calibration and uncertainty

Predicted probabilities should be meaningfully calibrated.

Select by accuracy and then evaluate calibration.

Robustness

Outputs should stay correct under input transformations—for example, local robustness (models trained on modern English applied to archaic English), and adversarial robustness (adversarial examples designed to mislead the model).