The Right Time Matters: Data Arrangement Affects Zero-Shot Generalization in Instruction Tuning

TL;DR

This paper investigates how data arrangement during instruction tuning influences zero-shot generalization in large language models, revealing that early training data exposure and similarity-based arrangements significantly enhance unseen task performance.

Contribution

It introduces a novel data arrangement framework, Test-centric Multi-turn Arrangement, demonstrating its effectiveness in improving zero-shot generalization and continual learning.

Findings

Zero-shot generalization occurs early during instruction tuning.

Timing and similarity of training data greatly impact generalization.

Proposed framework enhances continual learning and reduces loss.

Abstract

Understanding alignment techniques begins with comprehending zero-shot generalization brought by instruction tuning, but little of the mechanism has been understood. Existing work has largely been confined to the task level, without considering that tasks are artificially defined and, to LLMs, merely consist of tokens and representations. To bridge this gap, we investigate zero-shot generalization from the perspective of the data itself. We first demonstrate that zero-shot generalization happens very early during instruction tuning, with loss serving as a stable indicator. Next, we investigate training data arrangement through similarity and granularity perspectives, confirming that the timing of exposure to certain training examples may greatly facilitate generalization on unseen tasks. Finally, we propose a more grounded training data arrangement framework, Test-centric Multi-turn Arrangement, and show its effectiveness in promoting continual learning and further loss reduction. For the first time, we show that zero-shot generalization during instruction tuning is a form of similarity-based generalization between training and test data at the instance level. Our code is released at https://github.com/thunlp/Dynamics-of-Zero-Shot-Generalization.