You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
Low-rank adaptations (LoRA) are often employed to fine-tune large languagemodels (LLMs) for new tasks. This paper investigates LoRA composability forcross-task generalization and introduces LoraHub, a strategic framework devisedfor the purposive assembly of LoRA modules trained on diverse given tasks, withthe objective of achieving adaptable performance on unseen tasks. With just afew examples from a novel task, LoraHub enables the fluid combination ofmultiple LoRA modules, eradicating the need for human expertise. Notably, thecomposition requires neither additional model parameters nor gradients. Ourempirical results, derived from the Big-Bench Hard (BBH) benchmark, suggestthat LoraHub can effectively mimic the performance of in-context learning infew-shot scenarios, excluding the necessity of in-context examples alongsideeach inference input. A significant contribution of our research is thefostering of a community for LoRA, where users can share their trained LoRAmodules, thereby facilitating their application to new tasks. We anticipatethis resource will widen access to and spur advancements in generalintelligence as well as LLMs in production. Code will be available athttps://github.com/sail-sg/lorahub.
AkihikoWatanabe
changed the title
あ
LoraHub: Efficient Cross-Task Generalization via Dynamic LoRA
Composition, Chengsong Huang+, N/A, arXiv'23
Aug 8, 2023
URL
Affiliations
Abstract
Translation (by gpt-3.5-turbo)
Summary (by gpt-3.5-turbo)
The text was updated successfully, but these errors were encountered: