ResearchTrend.AI
  • Papers
  • Communities
  • Events
  • Blog
  • Pricing
Papers
Communities
Social Events
Terms and Conditions
Pricing
Parameter LabParameter LabTwitterGitHubLinkedInBlueskyYoutube

© 2025 ResearchTrend.AI, All rights reserved.

  1. Home
  2. Papers
  3. 2305.12567
22
2

Model-Generated Pretraining Signals Improves Zero-Shot Generalization of Text-to-Text Transformers

21 May 2023
Linyuan Gong
Chenyan Xiong
Xiaodong Liu
Payal Bajaj
Yiqing Xie
Alvin Cheung
Jianfeng Gao
Xia Song
    VLM
    AI4CE
ArXivPDFHTML
Abstract

This paper explores the effectiveness of model-generated signals in improving zero-shot generalization of text-to-text Transformers such as T5. We study various designs to pretrain T5 using an auxiliary model to construct more challenging token replacements for the main model to denoise. Key aspects under study include the decoding target, the location of the RTD head, and the masking pattern. Based on these studies, we develop a new model, METRO-T0, which is pretrained using the redesigned ELECTRA-Style pretraining strategies and then prompt-finetuned on a mixture of NLP tasks. METRO-T0 outperforms all similar-sized baselines on prompted NLP benchmarks, such as T0 Eval and MMLU, and rivals the state-of-the-art T0-11B model with only 8% of its parameters. Our analysis on model's neural activation and parameter sensitivity reveals that the effectiveness of METRO-T0 stems from more balanced contribution of parameters and better utilization of their capacity. The code and model checkpoints are available at https://github.com/gonglinyuan/metro_t0.

View on arXiv
Comments on this paper