Calibrate Before Use: Improving Few-Shot Performance of Language Models

TL;DR

This paper introduces a calibration method to stabilize and improve the few-shot learning performance of language models like GPT-3 by adjusting for their inherent answer biases, leading to significant accuracy gains.

Contribution

The authors propose a novel calibration technique that estimates and corrects language model biases, substantially enhancing few-shot task performance and consistency.

Findings

Calibration improves GPT-3 accuracy by up to 30%.

Reduces variability across different prompt choices.

Effective across diverse NLP tasks.

Abstract

GPT-3 can perform numerous tasks when provided a natural language prompt that contains a few training examples. We show that this type of few-shot learning can be unstable: the choice of prompt format, training examples, and even the order of the training examples can cause accuracy to vary from near chance to near state-of-the-art. We demonstrate that this instability arises from the bias of language models towards predicting certain answers, e.g., those that are placed near the end of the prompt or are common in the pre-training data. To mitigate this, we first estimate the model's bias towards each answer by asking for its prediction when given the training prompt and a content-free test input such as "N/A". We then fit calibration parameters that cause the prediction for this input to be uniform across answers. On a diverse set of tasks, this contextual calibration procedure substantially improves GPT-3 and GPT-2's average accuracy (up to 30.0% absolute) and reduces variance across different choices of the prompt.