Embedded Files
Fine-tuned LLM for Private LLM/RAG
Admin 2025-05-01 👁️ 207
We aim to investigate natural language processing methods that support the development of private, domain-adapted large language models. Our focus is on exploring and fine-tuning state-of-the-art algorithms to achieve:
Effective long-term memory integration (LLM/RAG) that allows models to recall, update, and manage knowledge in a private setting.
Improved performance in retrieval-augmented generation pipelines by adapting LLMs to specialized domains and sensitive data.
Privacy and security by design, ensuring that models can be deployed safely for research and applied systems where data confidentiality is critical.
This direction strengthens both the technical foundation (through fine-tuning and RAG improvements) and the practical trustworthiness of LLMs in private contexts.
1.Motivation
- Using thrid-party LLMs like ChatGPT gives the best performance on open domain tasks but gives rise to the risk of exposing sensitive client data
- Using an open LLM like LLaMA2 on a privately hosted server can solve this issue but results in lower performance
- Thus, there is a need to develop private LLMs with better performance on tasks requested by clients while maintaining privacy of client data
2. Research Goal and Issue
- Goal : Develop a private LLM/RAG(Retrieval-Augmented Generation) system
Develop finetuning methods using multitask leraning to improve the response generation ability of the private LLM
Improve the performance of the LLM by combining it with ISPL RAG model
- Issue
MoE models encompass parameter complexity and training instability
Hypernetworks : pose scalability challenges
3. Approach
- Develop a private LLM
Collect and process data for building training dataset
Finetune LLM using multitask learning
Integrate finetuned LLM with RAG
4.Result
- the performance evaluation of the fine-tuned model was much better than the pre-trained model
Page updated
Report abuse