SpacTor-T5: Pre-training T5 Models with Span Corruption and Replaced Token Detection

Fuente: arXiv
Salvato in:
Dettagli Bibliografici
Autori principali: Ye, Ke, Jiang, Heinrich, Rostamizadeh, Afshin, Chakrabarti, Ayan, DeSalvo, Giulia, Kagy, Jean-François, Karydas, Lazaros, Citovsky, Gui, Kumar, Sanjiv
Natura: Preprint
Pubblicazione: 2024
Soggetti:
Accesso online:
Tags: Aggiungi Tag
Nessun Tag, puoi essere il primo ad aggiungerne!!
_version_ 1866929221296717824
author Ye, Ke
Jiang, Heinrich
Rostamizadeh, Afshin
Chakrabarti, Ayan
DeSalvo, Giulia
Kagy, Jean-François
Karydas, Lazaros
Citovsky, Gui
Kumar, Sanjiv
author_facet Ye, Ke
Jiang, Heinrich
Rostamizadeh, Afshin
Chakrabarti, Ayan
DeSalvo, Giulia
Kagy, Jean-François
Karydas, Lazaros
Citovsky, Gui
Kumar, Sanjiv
contents Pre-training large language models is known to be extremely resource intensive and often times inefficient, under-utilizing the information encapsulated in the training text sequences. In this paper, we present SpacTor, a new training procedure consisting of (1) a hybrid objective combining span corruption (SC) and token replacement detection (RTD), and (2) a two-stage curriculum that optimizes the hybrid objective over the initial $τ$ iterations, then transitions to standard SC loss. We show empirically that the effectiveness of the hybrid objective is tied to the two-stage pre-training schedule, and provide extensive analysis on why this is the case. In our experiments with encoder-decoder architectures (T5) on a variety of NLP tasks, SpacTor-T5 yields the same downstream performance as standard SC pre-training, while enabling a 50% reduction in pre-training iterations and 40% reduction in total FLOPs. Alternatively, given the same amount of computing budget, we find that SpacTor results in significantly improved downstream benchmark performance.
format Preprint
id arxiv_https___arxiv_org_abs_2401_13160
institution arXiv
publishDate 2024
record_format arxiv
spellingShingle SpacTor-T5: Pre-training T5 Models with Span Corruption and Replaced Token Detection
Ye, Ke
Jiang, Heinrich
Rostamizadeh, Afshin
Chakrabarti, Ayan
DeSalvo, Giulia
Kagy, Jean-François
Karydas, Lazaros
Citovsky, Gui
Kumar, Sanjiv
Machine Learning
Computation and Language
Pre-training large language models is known to be extremely resource intensive and often times inefficient, under-utilizing the information encapsulated in the training text sequences. In this paper, we present SpacTor, a new training procedure consisting of (1) a hybrid objective combining span corruption (SC) and token replacement detection (RTD), and (2) a two-stage curriculum that optimizes the hybrid objective over the initial $τ$ iterations, then transitions to standard SC loss. We show empirically that the effectiveness of the hybrid objective is tied to the two-stage pre-training schedule, and provide extensive analysis on why this is the case. In our experiments with encoder-decoder architectures (T5) on a variety of NLP tasks, SpacTor-T5 yields the same downstream performance as standard SC pre-training, while enabling a 50% reduction in pre-training iterations and 40% reduction in total FLOPs. Alternatively, given the same amount of computing budget, we find that SpacTor results in significantly improved downstream benchmark performance.
title SpacTor-T5: Pre-training T5 Models with Span Corruption and Replaced Token Detection
topic Machine Learning
Computation and Language
url https://arxiv.org/abs/2401.13160