The T5 model has been a staple for NLP research for the last years. Both its size and its approach to formulate all NLP tasks as prompt-based language modeling make it a convenient choice to tackle new challenges and provides a strong baseline for most current datasets. ExT5 pushes T5 to its limits by pre-training not only on self-supervised mask filling, but also at the same time on 107 different supervised NLP tasks, which is their new ExMix dataset. The resulting model compares very favorably to T5 when fine-tuned to downstream tasks.
0:00 - Intro & Overview
2:15 - Recap: The T5 model
3:55 - The ExT5 model and task formulations
8:10 - ExMix dataset
9:35 - Do different tasks help each other?
16:50 - Which tasks should we include?
20:30 - Pre-Training vs Pre-Finetuning
23:00 - A few hypotheses about what’s going on
27:20 - How much self-supervised data to use?
34:15 - More experimental results
38:40 - Conclusion & Summary
Despite the recent success of multi-task learning and transfer learning for natural language processing (NLP), few works have systematically studied the effect of scaling up the number of tasks during pre-training. Towards this goal, this paper introduces ExMix (Extreme Mixture): a massive collection of 107 supervised NLP tasks across diverse domains and task-families. Using ExMix, we study the effect of multi-task pre-training at the largest scale to date, and analyze co-training transfer amongst common families of tasks. Through this analysis, we show that manually curating an ideal set of tasks for multi-task pre-training is not straightforward, and that multi-task scaling can vastly improve models on its own. Finally, we propose ExT5: a model pre-trained using a multi-task objective of self-supervised span denoising and supervised ExMix. Via extensive experiments, we show that ExT5 outperforms strong T5 baselines on SuperGLUE, GEM, Rainbow, Closed-Book QA tasks, and several tasks outside of ExMix. ExT5 also significantly improves sample efficiency while pre-training.
Authors: Vamsi Aribandi, Yi Tay, Tal Schuster, Jinfeng Rao, Huaixiu Steven Zheng, Sanket Vaibhav Mehta, Honglei Zhuang, Vinh Q. Tran, Dara Bahri, Jianmo Ni, Jai Gupta, Kai Hui, Sebastian Ruder, Donald Metzler