| description |
|---|
Low-Rank Adaptation of Large Language Models. It is a promising technique for adapting LLMs to specific tasks in a computationally efficient manner. |
LoRA is a low-rank approximation algorithm that uses SVD to reduce the dimensionality of a dataset. It works by first computing the SVD of the data matrix. Then, it selects a subset of the singular values and corresponding singular vectors to form a low-rank approximation of the data matrix.
LoRA is a technique for adapting large language models (LLMs) to specific tasks. LoRA reduces the number of trainable parameters in an LLM by learning pairs of rank-decomposition matrices while freezing the original weights. This vastly reduces the storage requirement for LLMs adapted to specific tasks and enables efficient task-switching during deployment all without introducing inference latency. LoRA also outperforms several other adaptation methods including adapter, prefix-tuning, and fine-tuning.
It works by first pre-training an LLM on a large corpus of text. Once the LLM is pre-trained, LoRA uses a technique called rank decomposition to learn a set of low-rank matrices that can be used to adapt the LLM to a specific task.
The rank decomposition process is performed by first computing the singular value decomposition (SVD) of the LLM's weights. The SVD decomposes the weights into a set of left singular vectors, a set of right singular vectors, and a diagonal matrix of singular values. The left and right singular vectors are then used to construct the low-rank matrices.
Once the low-rank matrices are constructed, they are used to adapt the LLM to a specific task. This is done by multiplying the LLM's weights by the low-rank matrices. The multiplication of the weights by the low-rank matrices has the effect of reducing the number of trainable parameters in the LLM.