Rebuilding ROME : Resolving Model Collapse during Sequential Model Editing
Fuente:
arXiv
Salvato in:
| Autori principali: | , , |
|---|---|
| Natura: | Preprint |
| Pubblicazione: |
2024
|
| Soggetti: | |
| Accesso online: | |
| Tags: |
Aggiungi Tag
Nessun Tag, puoi essere il primo ad aggiungerne!!
|
| _version_ | 1866910639832694784 |
|---|---|
| author | Gupta, Akshat Baskaran, Sidharth Anumanchipalli, Gopala |
| author_facet | Gupta, Akshat Baskaran, Sidharth Anumanchipalli, Gopala |
| contents | Recent work using Rank-One Model Editing (ROME), a popular model editing method, has shown that there are certain facts that the algorithm is unable to edit without breaking the model. Such edits have previously been called disabling edits. These disabling edits cause immediate model collapse and limits the use of ROME for sequential editing. In this paper, we show that disabling edits are an artifact of irregularities in the implementation of ROME. With this paper, we provide a more stable implementation ROME, which we call r-ROME and show that model collapse is no longer observed when making large scale sequential edits with r-ROME, while further improving generalization and locality of model editing compared to the original implementation of ROME. We also provide a detailed mathematical explanation of the reason behind disabling edits. |
| format | Preprint |
| id |
arxiv_https___arxiv_org_abs_2403_07175 |
| institution | arXiv |
| publishDate | 2024 |
| record_format | arxiv |
| spellingShingle | Rebuilding ROME : Resolving Model Collapse during Sequential Model Editing Gupta, Akshat Baskaran, Sidharth Anumanchipalli, Gopala Computation and Language Artificial Intelligence Recent work using Rank-One Model Editing (ROME), a popular model editing method, has shown that there are certain facts that the algorithm is unable to edit without breaking the model. Such edits have previously been called disabling edits. These disabling edits cause immediate model collapse and limits the use of ROME for sequential editing. In this paper, we show that disabling edits are an artifact of irregularities in the implementation of ROME. With this paper, we provide a more stable implementation ROME, which we call r-ROME and show that model collapse is no longer observed when making large scale sequential edits with r-ROME, while further improving generalization and locality of model editing compared to the original implementation of ROME. We also provide a detailed mathematical explanation of the reason behind disabling edits. |
| title | Rebuilding ROME : Resolving Model Collapse during Sequential Model Editing |
| topic | Computation and Language Artificial Intelligence |
| url | https://arxiv.org/abs/2403.07175 |