FLAN-T5

Google · October 2022

activeOpen Sourceencoder decodertext

Why It Matters

Showed that instruction tuning at scale could make smaller models competitive with much larger ones, influencing the development of efficient fine-tuning techniques across the industry.

Description

Google's instruction-tuned version of T5, fine-tuned on over 1,800 tasks described via natural language instructions. Demonstrated that instruction tuning dramatically improves zero-shot and few-shot performance across virtually all NLP tasks.

Key Innovations

Instruction Tuning
Instruction TuningFine-tuning a model on instruction-response pairs so it follows user commands more reliably.
Few-Shot
Few-ShotLearning from just a handful of examples provided in the prompt, without retraining.

Family Tree

Built On

Lineage

BERTT5FLAN-T5

External Links