Tools

LEADER 03443cam a2200517 i 4500

001 on1450451766;

003 OCoLC;

005 20260527050256.0;

006 m o d ;

007 cr unu||||||||;

008 240731s2024 ncua obm 000 0 eng d;

035 a| (OCoLC)1450451766;

090 a| Q335;

049 a| EREE;

100 1 a| Stein, Gabrielle,e| author.?| UNAUTHORIZED;

245 14 a| The perils of generative model inbreeding :b| evaluating the consequences of cross-model training in large language models /c| by Gabrielle Stein.;

264 1 a| [Greenville, N.C.] :b| [East Carolina University],c| 2024.;

300 a| 1 online resource (44 pages) :b| illustrations (chiefly color);

336 a| textb| txt2| rdacontent;

337 a| computerb| c2| rdamedia;

338 a| online resourceb| cr2| rdacarrier;

347 a| text fileb| application/pdfc| 491.42 KB2| rda;

538 a| System requirements: Adobe Reader.;

538 a| Mode of access: World Wide Web.;

502 b| M.S.c| East Carolina Universityd| 2024.;

500 a| Presented to the Faculty of the Department of Computer Science;

500 a| Advisor: Nic Herndon;

500 a| Title from PDF t.p. (viewed August 12, 2025).;

520 3 a| What happens when the output of generative AI models is included in the training data of new models? With the rise of generative AI content online, and considering that most training data for AI models is sourced from the Internet, concerns have arisen about how this generated content might taint future training datasets. Existing research has evaluated the effect of models consuming their own output, and has shown that the output of self-consuming models degrades with each successive generation of re-training, a phenomenon coined as "model collapse.'' This degradation takes the form of a loss of diversity in the output of the model. Currently there is limited research on the impact of models consuming other models' output, specifically large language models. In this study we aimed to determine the effect of training a model on a different model's output. Additionally, we developed a potential solution to prevent "model collapse.'' Guaranteeing the majority of training data is guaranteed to be human-generated (non-synthetic) data has been shown to mitigate the loss of diversity caused by "model collapse.'' Given that AI models are here to stay, the methods for developing new models will need to evolve to address this issue, ensuring that AI development can continue to progress and improve.;

504 a| Includes bibliographical references.;

650 0 a| Artificial intelligencex| Evaluation.=| ^A15394;

650 0 a| Machine learningx| Evaluation.=| ^A145211;

650 0 a| Natural language processing (Computer science)=| ^A235570;

653 a| AI;

653 a| Large language model;

653 a| model collapse;

653 a| generative AI;

700 1 a| Herndon, Nic,e| degree supervisor.?| UNAUTHORIZED;

710 2 a| East Carolina University.b| Department of Computer Science.?| UNAUTHORIZED;

856 40 z| Access via ScholarShipu| http://hdl.handle.net/10342/13448;

994 a| C0b| ERE;

596 a| 1 4;

998 a| 6462771;

999 a| CLICK ON WEB ADDRESSw| ASISc| 1i| 6462771-1001l| JNETm| JOYNERr| Ys| Yt| JNE3ETDu| 7/31/2024x| ETDz| JERESOURCE;

999 a| CLICK ON WEB ADDRESSw| ASISc| 1i| 6462771-2001l| HSLELECm| HSLr| Ys| Yt| HEETDu| 7/31/2024x| ETDz| HERESOURCE;

The perils of generative model inbreeding : evaluating the consequences of cross-model training in large language models / by Gabrielle Stein.

Author/creator	Stein, Gabrielle author.
Other author	Herndon, Nic, degree supervisor.
Other author	East Carolina University. Department of Computer Science.
Format	Theses and dissertations
Publication	[Greenville, N.C.] : [East Carolina University], 2024.
Description	1 online resource (44 pages) : illustrations (chiefly color)
Supplemental Content	Access via ScholarShip
Subjects	Artificial intelligence--Evaluation. Machine learning--Evaluation. Natural language processing (Computer science)

Summary	What happens when the output of generative AI models is included in the training data of new models? With the rise of generative AI content online, and considering that most training data for AI models is sourced from the Internet, concerns have arisen about how this generated content might taint future training datasets. Existing research has evaluated the effect of models consuming their own output, and has shown that the output of self-consuming models degrades with each successive generation of re-training, a phenomenon coined as "model collapse.'' This degradation takes the form of a loss of diversity in the output of the model. Currently there is limited research on the impact of models consuming other models' output, specifically large language models. In this study we aimed to determine the effect of training a model on a different model's output. Additionally, we developed a potential solution to prevent "model collapse.'' Guaranteeing the majority of training data is guaranteed to be human-generated (non-synthetic) data has been shown to mitigate the loss of diversity caused by "model collapse.'' Given that AI models are here to stay, the methods for developing new models will need to evolve to address this issue, ensuring that AI development can continue to progress and improve.
General note	Presented to the Faculty of the Department of Computer Science
General note	Advisor: Nic Herndon
General note	Title from PDF t.p. (viewed August 12, 2025).
Dissertation note	M.S. East Carolina University 2024.
Bibliography note	Includes bibliographical references.
Technical details	System requirements: Adobe Reader.
Technical details	Mode of access: World Wide Web.

Availability

Library	Location	Call Number	Status	Item Actions
	Electronic Resources	Access Content Online	✔ Available

The perils of generative model inbreeding : evaluating the consequences of cross-model training in large language models / by Gabrielle Stein.

Click here for more information about this title

Availability