Authors :
M Krishna Satya Varma; Koteswara Rao; Sai Ganesh; Venkat Sai Koushik; Rama Krishnam Raju
Volume/Issue :
Volume 9 - 2024, Issue 4 - April
Google Scholar :
https://tinyurl.com/5n7w688p
Scribd :
https://tinyurl.com/2mpu2r5n
DOI :
https://doi.org/10.38124/ijisrt/IJISRT24APR285
Abstract :
Despite their ability to store information and
excel at many NLP tasks with fine-tuning, large language
models tend to have issues about accurately accessing and
altering knowledge, which leads to performance gaps in
knowledge-intensive tasks compared to domain-specific
architectures. Additionally, these models face problems
when it comes to having transparent decision-making
processes or updating their world knowledge. To mitigate
these limitations, we propose a Retrieval Augmented
Generation (RAG) system by improving the Mistral7B
model specifically for RAG tasks. The novel training
technique includes Parameter-Efficient Fine-Tuning
(PEFT) which enables efficient adaptation of large pre-
trained models on-the-fly according to task-specific
requirements while reducing computational costs. In
addition, this system combines pre-trained embedding
models that use pre-trained cross-encoders for effective
retrieval and reranking of information. This RAG system
will thus leverage these state-of-the-art methodologies
towards achieving top performances in a range of NLP
tasks such as question answering and summarization.
Keywords :
Component: RAG, PEFT, Cross Encoders.
Despite their ability to store information and
excel at many NLP tasks with fine-tuning, large language
models tend to have issues about accurately accessing and
altering knowledge, which leads to performance gaps in
knowledge-intensive tasks compared to domain-specific
architectures. Additionally, these models face problems
when it comes to having transparent decision-making
processes or updating their world knowledge. To mitigate
these limitations, we propose a Retrieval Augmented
Generation (RAG) system by improving the Mistral7B
model specifically for RAG tasks. The novel training
technique includes Parameter-Efficient Fine-Tuning
(PEFT) which enables efficient adaptation of large pre-
trained models on-the-fly according to task-specific
requirements while reducing computational costs. In
addition, this system combines pre-trained embedding
models that use pre-trained cross-encoders for effective
retrieval and reranking of information. This RAG system
will thus leverage these state-of-the-art methodologies
towards achieving top performances in a range of NLP
tasks such as question answering and summarization.
Keywords :
Component: RAG, PEFT, Cross Encoders.
