Analyzing constrained LLM through PDFA-learning

Carrasco, Matías - Mayr, Franz - Yovine, Sergio - Kidd, Johny - Iturbide, Martín - da Silva, Juan - Garat, Alejo

Resumen:

We define a congruence that copes with null next-symbol probabilities that arise when the output of a language model is constrained by some means during text generation. We develop an algorithm for efficiently learning the quotient with respect to this congruence and evaluate it on case studies for analyzing statistical properties of LLM.


Detalles Bibliográficos
2024
Agencia Nacional de Investigación e Innovación
Inteligencia artificial generativa
Grandes modelos de lenguaje
Ciencias Naturales y Exactas
Ciencias de la Computación e Información
Ciencias de la Computación
Inglés
Agencia Nacional de Investigación e Innovación
REDI
https://hdl.handle.net/20.500.12381/3624
Acceso abierto
Reconocimiento 4.0 Internacional. (CC BY)
_version_ 1814959263692357632
author Carrasco, Matías
author2 Mayr, Franz
Yovine, Sergio
Kidd, Johny
Iturbide, Martín
da Silva, Juan
Garat, Alejo
author2_role author
author
author
author
author
author
author_facet Carrasco, Matías
Mayr, Franz
Yovine, Sergio
Kidd, Johny
Iturbide, Martín
da Silva, Juan
Garat, Alejo
author_role author
bitstream.checksum.fl_str_mv a4ce09f01b5dd771727aa05c73851623
3e66590e86fd8cb1fba541387c99c839
bitstream.checksumAlgorithm.fl_str_mv MD5
MD5
bitstream.url.fl_str_mv https://redi.anii.org.uy/jspui/bitstream/20.500.12381/3624/2/license.txt
https://redi.anii.org.uy/jspui/bitstream/20.500.12381/3624/1/2406.08269v2.pdf
collection REDI
dc.creator.none.fl_str_mv Carrasco, Matías
Mayr, Franz
Yovine, Sergio
Kidd, Johny
Iturbide, Martín
da Silva, Juan
Garat, Alejo
dc.date.accessioned.none.fl_str_mv 2024-09-12T11:00:26Z
dc.date.available.none.fl_str_mv 2024-09-12T11:00:26Z
dc.date.issued.none.fl_str_mv 2024-07-07
dc.description.abstract.none.fl_txt_mv We define a congruence that copes with null next-symbol probabilities that arise when the output of a language model is constrained by some means during text generation. We develop an algorithm for efficiently learning the quotient with respect to this congruence and evaluate it on case studies for analyzing statistical properties of LLM.
dc.description.sponsorship.none.fl_txt_mv Agencia Nacional de Investigación e Innovación
dc.identifier.anii.es.fl_str_mv IA_1_2022_1_173516
FMV_1_2023_1_175864
dc.identifier.uri.none.fl_str_mv https://hdl.handle.net/20.500.12381/3624
dc.language.iso.none.fl_str_mv eng
dc.relation.es.fl_str_mv https://arxiv.org/abs/2406.08269
dc.rights.*.fl_str_mv Acceso abierto
dc.rights.license.none.fl_str_mv Reconocimiento 4.0 Internacional. (CC BY)
dc.rights.none.fl_str_mv info:eu-repo/semantics/openAccess
dc.source.none.fl_str_mv reponame:REDI
instname:Agencia Nacional de Investigación e Innovación
instacron:Agencia Nacional de Investigación e Innovación
dc.subject.anii.none.fl_str_mv Ciencias Naturales y Exactas
Ciencias de la Computación e Información
Ciencias de la Computación
dc.subject.es.fl_str_mv Inteligencia artificial generativa
Grandes modelos de lenguaje
dc.title.none.fl_str_mv Analyzing constrained LLM through PDFA-learning
dc.type.es.fl_str_mv Preprint
dc.type.none.fl_str_mv info:eu-repo/semantics/preprint
description We define a congruence that copes with null next-symbol probabilities that arise when the output of a language model is constrained by some means during text generation. We develop an algorithm for efficiently learning the quotient with respect to this congruence and evaluate it on case studies for analyzing statistical properties of LLM.
eu_rights_str_mv openAccess
format preprint
id REDI_a1278470b6c20ec1be1c722d320323a9
identifier_str_mv IA_1_2022_1_173516
FMV_1_2023_1_175864
instacron_str Agencia Nacional de Investigación e Innovación
institution Agencia Nacional de Investigación e Innovación
instname_str Agencia Nacional de Investigación e Innovación
language eng
network_acronym_str REDI
network_name_str REDI
oai_identifier_str oai:redi.anii.org.uy:20.500.12381/3624
publishDate 2024
reponame_str REDI
repository.mail.fl_str_mv jmaldini@anii.org.uy
repository.name.fl_str_mv REDI - Agencia Nacional de Investigación e Innovación
repository_id_str 9421
rights_invalid_str_mv Reconocimiento 4.0 Internacional. (CC BY)
Acceso abierto
spelling Reconocimiento 4.0 Internacional. (CC BY)Acceso abiertoinfo:eu-repo/semantics/openAccess2024-09-12T11:00:26Z2024-09-12T11:00:26Z2024-07-07https://hdl.handle.net/20.500.12381/3624IA_1_2022_1_173516FMV_1_2023_1_175864We define a congruence that copes with null next-symbol probabilities that arise when the output of a language model is constrained by some means during text generation. We develop an algorithm for efficiently learning the quotient with respect to this congruence and evaluate it on case studies for analyzing statistical properties of LLM.Agencia Nacional de Investigación e Innovaciónenghttps://arxiv.org/abs/2406.08269Inteligencia artificial generativaGrandes modelos de lenguajeCiencias Naturales y ExactasCiencias de la Computación e InformaciónCiencias de la ComputaciónAnalyzing constrained LLM through PDFA-learningPreprintinfo:eu-repo/semantics/preprintUniversidad ORT Uruguay//Ciencias Naturales y Exactas/Ciencias de la Computación e Información/Ciencias de la Computaciónreponame:REDIinstname:Agencia Nacional de Investigación e Innovacióninstacron:Agencia Nacional de Investigación e InnovaciónCarrasco, MatíasMayr, FranzYovine, SergioKidd, JohnyIturbide, Martínda Silva, JuanGarat, AlejoLICENSElicense.txtlicense.txttext/plain; charset=utf-84967https://redi.anii.org.uy/jspui/bitstream/20.500.12381/3624/2/license.txta4ce09f01b5dd771727aa05c73851623MD52ORIGINAL2406.08269v2.pdf2406.08269v2.pdfapplication/pdf593640https://redi.anii.org.uy/jspui/bitstream/20.500.12381/3624/1/2406.08269v2.pdf3e66590e86fd8cb1fba541387c99c839MD5120.500.12381/36242024-09-12 08:00:28.001oai:redi.anii.org.uy:20.500.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Gobiernohttps://www.anii.org.uy/https://redi.anii.org.uy/oai/requestjmaldini@anii.org.uyUruguayopendoar:94212024-09-12T11:00:28REDI - Agencia Nacional de Investigación e Innovaciónfalse
spellingShingle Analyzing constrained LLM through PDFA-learning
Carrasco, Matías
Inteligencia artificial generativa
Grandes modelos de lenguaje
Ciencias Naturales y Exactas
Ciencias de la Computación e Información
Ciencias de la Computación
title Analyzing constrained LLM through PDFA-learning
title_full Analyzing constrained LLM through PDFA-learning
title_fullStr Analyzing constrained LLM through PDFA-learning
title_full_unstemmed Analyzing constrained LLM through PDFA-learning
title_short Analyzing constrained LLM through PDFA-learning
title_sort Analyzing constrained LLM through PDFA-learning
topic Inteligencia artificial generativa
Grandes modelos de lenguaje
Ciencias Naturales y Exactas
Ciencias de la Computación e Información
Ciencias de la Computación
url https://hdl.handle.net/20.500.12381/3624

Resultados similares