LLM Finetuning with RAG (Retrieval Augmented Generation)

Enterprises look into capitalizing on the unquestionable qualitative change in generative AI. The central problem of manipulating or combining domain-specific knowledge or local data with the foundation #LLM pipelines is significant. The solutions generally rely on "finetuning" procedures. A few leading methods include transfer learning, #BYOM, and several sequential architectures. #RAG (retrieval augmented generation) is a way to improve LLMs. This paper presents excellent results based on standard evaluation procedures.

A pre-trained document retriever (DPR), used by the authors, is a query understanding and retrieval procedure. A pre-trained #seq2seq encoder/decoder processes the context, which is typically containing the original query and the retrieved (context) documents. The encoder/decoder is based on #BART in this paper.

The parametric memory algorithm that uses a pre-trained "sequence to sequence" #transformer is well-documented in the ML literature. A non-parametric memory component explicitly contains vectorized content from large corpora. Sometimes those representations are simply called document indexes. The characteristics of the solution that relies on non-parametric memory are huge. There is no re-training necessary if we change the source of knowledge. On the other hand, BART or other parametric models need to be re-trained if we need to unfreeze the knowledge base. #ai #training #learning #finetuning #LLM #llms #paremetricmemory #ml

https://arxiv.org/pdf/2005.11401.pdf