Distilling Instruction-following Abilities of Large Language Models with Task-aware Curriculum Planning

TL;DR

This paper introduces TAPIR, a curriculum-based distillation method that improves large language models' instruction-following abilities by systematically selecting and balancing task difficulties, leading to superior performance with less data.

Contribution

The paper proposes TAPIR, a novel multi-round distillation framework that incorporates task-aware curriculum planning and automatic response refinement to enhance LLM instruction tuning.

Findings

Student LLMs outperform larger models on benchmarks.

TAPIR achieves better results with less training data.

Curriculum planning systematically improves instruction-following abilities.

Abstract

Instruction tuning aims to align large language models (LLMs) with open-domain instructions and human-preferred responses. While several studies have explored autonomous approaches to distilling and annotating instructions from powerful proprietary LLMs, such as ChatGPT, they often neglect the impact of the distributions and characteristics of tasks, together with the varying difficulty of instructions in training sets. This oversight can lead to imbalanced knowledge capabilities and poor generalization powers of student LLMs. To address these challenges, we introduce Task-Aware Curriculum Planning for Instruction Refinement (TAPIR), a multi-round distillation framework that utilizes an oracle LLM to select instructions that are difficult for a student LLM to follow. To balance the student's capabilities, task distributions in training sets are adjusted with responses automatically refined according to their corresponding tasks. In addition, by incorporating curriculum planning, our approach systematically escalates the difficulty levels of tasks, progressively enhancing the student LLM's capabilities. We rigorously evaluate TAPIR using several widely recognized benchmarks (such as AlpacaEval 2.0, MT-Bench, etc.) and multiple student LLMs. Empirical results demonstrate that student LLMs, trained with our method and less training data, outperform larger instruction-tuned models and strong distillation baselines.