The Flan Collection: Designing Data and Methods for Effective Instruction Tuning

TL;DR

This paper analyzes the design choices behind instruction tuning methods, especially Flan 2022, highlighting the importance of task balancing, mixed prompt training, and demonstrating that instruction-tuned models like Flan-T5 are more efficient and effective for downstream tasks.

Contribution

The paper provides a detailed ablation study of Flan 2022's design decisions, revealing critical factors for successful instruction tuning and offering a publicly available dataset collection.

Findings

Task balancing and enrichment are crucial for effective instruction tuning.

Training with mixed prompt settings improves performance across evaluation scenarios.

Flan-T5 requires less finetuning and converges faster than T5 on downstream tasks.

Abstract

We study the design decisions of publicly available instruction tuning methods, and break down the development of Flan 2022 (Chung et al., 2022). Through careful ablation studies on the Flan Collection of tasks and methods, we tease apart the effect of design decisions which enable Flan-T5 to outperform prior work by 3-17%+ across evaluation settings. We find task balancing and enrichment techniques are overlooked but critical to effective instruction tuning, and in particular, training with mixed prompt settings (zero-shot, few-shot, and chain-of-thought) actually yields stronger (2%+) performance in all settings. In further experiments, we show Flan-T5 requires less finetuning to converge higher and faster than T5 on single downstream tasks, motivating instruction-tuned models as more computationally-efficient starting checkpoints for new tasks. Finally, to accelerate research on instruction tuning, we make the Flan 2022 collection of datasets, templates, and methods publicly available at https://github.com/google-research/FLAN/tree/main/flan/v2.