@inproceedings{krishna-etal-2023-downstream,
title = "Downstream Datasets Make Surprisingly Good Pretraining Corpora",
author = "Krishna, Kundan and
Garg, Saurabh and
Bigham, Jeffrey and
Lipton, Zachary",
editor = "Rogers, Anna and
Boyd-Graber, Jordan and
Okazaki, Naoaki",
booktitle = "Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers)",
month = jul,
year = "2023",
address = "Toronto, Canada",
publisher = "Association for Computational Linguistics",
url = "https://aclanthology.org/2023.acl-long.682",
doi = "10.18653/v1/2023.acl-long.682",
pages = "12207--12222",
abstract = "For most natural language processing tasks, the dominant practice is to finetune large pretrained transformer models (e.g., BERT) using smaller downstream datasets. Despite the success of this approach, it remains unclear to what extent these gainsare attributable to the massive background corpora employed for pretraining versus to the pretraining objectives themselves. This paper introduces a large-scale study of self-pretraining, where the same (downstream) training data is used for both pretraining and finetuning.In experiments addressing both ELECTRA and RoBERTa models and 10 distinct downstream classification datasets, we observe that self-pretraining rivals standard pretraining on the BookWiki corpus (despite using around 10x{--}500x less data), outperforming the latter on 7 and 5 datasets, respectively. Surprisingly, these task-specific pretrained models often perform well on other tasks,including the GLUE benchmark. Besides classification tasks, self-pretraining also provides benefits on structured output prediction tasks such as span based question answering and commonsense inference, often providing more than 50{\%} of the performance boosts provided by pretraining on the BookWiki corpus. Our results hint that in many scenarios, performance gains attributable to pretraining are driven primarily by the pretraining objective itself and are not always attributable to the use of external pretraining data in massive amounts. These findings are especially relevant in light of concerns about intellectual property and offensive content in web-scale pretraining data.",
}
<?xml version="1.0" encoding="UTF-8"?>
<modsCollection xmlns="http://www.loc.gov/mods/v3">
<mods ID="krishna-etal-2023-downstream">
<titleInfo>
<title>Downstream Datasets Make Surprisingly Good Pretraining Corpora</title>
</titleInfo>
<name type="personal">
<namePart type="given">Kundan</namePart>
<namePart type="family">Krishna</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Saurabh</namePart>
<namePart type="family">Garg</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Jeffrey</namePart>
<namePart type="family">Bigham</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Zachary</namePart>
<namePart type="family">Lipton</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<originInfo>
<dateIssued>2023-07</dateIssued>
</originInfo>
<typeOfResource>text</typeOfResource>
<relatedItem type="host">
<titleInfo>
<title>Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers)</title>
</titleInfo>
<name type="personal">
<namePart type="given">Anna</namePart>
<namePart type="family">Rogers</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Jordan</namePart>
<namePart type="family">Boyd-Graber</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Naoaki</namePart>
<namePart type="family">Okazaki</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<originInfo>
<publisher>Association for Computational Linguistics</publisher>
<place>
<placeTerm type="text">Toronto, Canada</placeTerm>
</place>
</originInfo>
<genre authority="marcgt">conference publication</genre>
</relatedItem>
<abstract>For most natural language processing tasks, the dominant practice is to finetune large pretrained transformer models (e.g., BERT) using smaller downstream datasets. Despite the success of this approach, it remains unclear to what extent these gainsare attributable to the massive background corpora employed for pretraining versus to the pretraining objectives themselves. This paper introduces a large-scale study of self-pretraining, where the same (downstream) training data is used for both pretraining and finetuning.In experiments addressing both ELECTRA and RoBERTa models and 10 distinct downstream classification datasets, we observe that self-pretraining rivals standard pretraining on the BookWiki corpus (despite using around 10x–500x less data), outperforming the latter on 7 and 5 datasets, respectively. Surprisingly, these task-specific pretrained models often perform well on other tasks,including the GLUE benchmark. Besides classification tasks, self-pretraining also provides benefits on structured output prediction tasks such as span based question answering and commonsense inference, often providing more than 50% of the performance boosts provided by pretraining on the BookWiki corpus. Our results hint that in many scenarios, performance gains attributable to pretraining are driven primarily by the pretraining objective itself and are not always attributable to the use of external pretraining data in massive amounts. These findings are especially relevant in light of concerns about intellectual property and offensive content in web-scale pretraining data.</abstract>
<identifier type="citekey">krishna-etal-2023-downstream</identifier>
<identifier type="doi">10.18653/v1/2023.acl-long.682</identifier>
<location>
<url>https://aclanthology.org/2023.acl-long.682</url>
</location>
<part>
<date>2023-07</date>
<extent unit="page">
<start>12207</start>
<end>12222</end>
</extent>
</part>
</mods>
</modsCollection>
%0 Conference Proceedings
%T Downstream Datasets Make Surprisingly Good Pretraining Corpora
%A Krishna, Kundan
%A Garg, Saurabh
%A Bigham, Jeffrey
%A Lipton, Zachary
%Y Rogers, Anna
%Y Boyd-Graber, Jordan
%Y Okazaki, Naoaki
%S Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers)
%D 2023
%8 July
%I Association for Computational Linguistics
%C Toronto, Canada
%F krishna-etal-2023-downstream
%X For most natural language processing tasks, the dominant practice is to finetune large pretrained transformer models (e.g., BERT) using smaller downstream datasets. Despite the success of this approach, it remains unclear to what extent these gainsare attributable to the massive background corpora employed for pretraining versus to the pretraining objectives themselves. This paper introduces a large-scale study of self-pretraining, where the same (downstream) training data is used for both pretraining and finetuning.In experiments addressing both ELECTRA and RoBERTa models and 10 distinct downstream classification datasets, we observe that self-pretraining rivals standard pretraining on the BookWiki corpus (despite using around 10x–500x less data), outperforming the latter on 7 and 5 datasets, respectively. Surprisingly, these task-specific pretrained models often perform well on other tasks,including the GLUE benchmark. Besides classification tasks, self-pretraining also provides benefits on structured output prediction tasks such as span based question answering and commonsense inference, often providing more than 50% of the performance boosts provided by pretraining on the BookWiki corpus. Our results hint that in many scenarios, performance gains attributable to pretraining are driven primarily by the pretraining objective itself and are not always attributable to the use of external pretraining data in massive amounts. These findings are especially relevant in light of concerns about intellectual property and offensive content in web-scale pretraining data.
%R 10.18653/v1/2023.acl-long.682
%U https://aclanthology.org/2023.acl-long.682
%U https://doi.org/10.18653/v1/2023.acl-long.682
%P 12207-12222
Markdown (Informal)
[Downstream Datasets Make Surprisingly Good Pretraining Corpora](https://aclanthology.org/2023.acl-long.682) (Krishna et al., ACL 2023)
ACL
- Kundan Krishna, Saurabh Garg, Jeffrey Bigham, and Zachary Lipton. 2023. Downstream Datasets Make Surprisingly Good Pretraining Corpora. In Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 12207–12222, Toronto, Canada. Association for Computational Linguistics.