자료유형  

 학위논문

Control Number  

0017164895

International Standard Book Number  

9798346381501

Dewey Decimal Classification Number  

004.678

Main Entry-Personal Name  

Taori, Rohan Rajiv.

Publication, Distribution, etc. (Imprint  

[S.l.] : Stanford University., 2024

Publication, Distribution, etc. (Imprint  

Ann Arbor : ProQuest Dissertations & Theses, 2024

Physical Description  

168 p.

General Note  

Source: Dissertations Abstracts International, Volume: 86-05, Section: A.

General Note  

Advisor: Hashimoto, Tatsunori.

Dissertation Note  

Thesis (Ph.D.)--Stanford University, 2024.

Summary, Etc.  

요약As large language models have grown in capability, their development pipelines have correspondingly grown in complexity.This complexity presents three key challenges: data pipelines that heavily utilize crowdworkers make reproducible comparisons difficult, increasingly general models expose a lack of reliable open-ended evaluation, and online data generated by deployed models distorts training distributions for future models trained on the internet. These challenges collectively impede rapid experimentation and deeper scientific understanding of AI systems that are being deployed widely.This thesis presents novel techniques and frameworks to address these challenges. To make data experimentation more accessible, we introduce a set of techniques to simulate crowdworker feedback with existing language models. We validate the quality of the simulated feedback by testing a variety of different learning algorithms, finding that method rankings between simulated feedback and real human feedback track closely. To make evaluation more reliable, we introduce an automated evaluation framework for open-ended generation with instruction-following language models. Finally, to help guide the deployment of models over time, we formalize and analyze systems where model interactions are recorded and used as future training data. We examine their stability over time and propose interventions to mitigate potential instabilities.These contributions formalize new ways to make language model training accessible with fewer resources, promoting more robust and careful study.

Subject Added Entry-Topical Term  

Internet.

Subject Added Entry-Topical Term  

Toxicity.

Subject Added Entry-Topical Term  

Gender.

Subject Added Entry-Topical Term  

Labeling.

Subject Added Entry-Topical Term  

Large language models.

Subject Added Entry-Topical Term  

Chatbots.

Subject Added Entry-Topical Term  

Sustainability.

Subject Added Entry-Topical Term  

Toxicology.

Added Entry-Corporate Name  

Stanford University.

Host Item Entry  

Dissertations Abstracts International. 86-05A.

Electronic Location and Access  

로그인을 한후 보실 수 있는 자료입니다.

Control Number  

joongbu:654274