Democratizing Advanced Attribution Analyses of Generative Language Models with the Inseq Toolkit
Uložené v:
| Názov: | Democratizing Advanced Attribution Analyses of Generative Language Models with the Inseq Toolkit |
|---|---|
| Autori: | Sarti, Gabriele, Feldhus, Nils, Qi, Jirui, Nissim, Malvina, Bisazza, Arianna |
| Informácie o vydavateľovi: | CEUR Workshop Proceedings (CEUR-WS.org), 2024. |
| Rok vydania: | 2024 |
| Predmety: | Feature Attribution, Generative Language Models, Python Toolkit, Natural Language Processing |
| Popis: | Inseq1 is a recent toolkit providing an intuitive and optimized interface to conduct feature attribution analyses of generative language models. In this work, we present the latest improvements to the library, including efforts to simplify the attribution of large language models on consumer hardware, additional attribution approaches, and a new client command to detect and attribute context usage in language model generations. We showcase an online demo using Inseq as an attribution backbone for context reliance analysis, and we highlight interesting contextual patterns in language model generations. Ultimately, this release furthers Inseq’s mission of centralizing good interpretability practices and enabling fair and reproducible model evaluations. |
| Druh dokumentu: | Conference object |
| Jazyk: | English |
| Prístupová URL adresa: | https://research.rug.nl/en/publications/f719d93e-ca37-4965-b935-69bc53a48a4f https://hdl.handle.net/11370/f719d93e-ca37-4965-b935-69bc53a48a4f |
| Rights: | CC BY |
| Prístupové číslo: | edsair.dris...01423..3f1b08ba700891a86a4a076231d8432f |
| Databáza: | OpenAIRE |
Nájsť tento článok vo Web of Science | Abstrakt: | Inseq1 is a recent toolkit providing an intuitive and optimized interface to conduct feature attribution analyses of generative language models. In this work, we present the latest improvements to the library, including efforts to simplify the attribution of large language models on consumer hardware, additional attribution approaches, and a new client command to detect and attribute context usage in language model generations. We showcase an online demo using Inseq as an attribution backbone for context reliance analysis, and we highlight interesting contextual patterns in language model generations. Ultimately, this release furthers Inseq’s mission of centralizing good interpretability practices and enabling fair and reproducible model evaluations. |
|---|