Efficient Fine-Tuning of Large Language Models

Efficient Fine-Tuning of Large Language Models (LLMs) for Multiple Use Cases using Low-Rank Adaptation (LoRA)

Description:
Large Language Models (LLMs) such as LLaMA or GPT demonstrate remarkable potential across a wide range of industrial applications. However, training or fine-tuning such models is extremely resource-intensive and poses major challenges when hardware capacity is limited (e.g., restricted GPU memory). One promising approach to overcome these limitations is Low-Rank Adaptation (LoRA).

Within this project, students will investigate whether it is possible to maintain a single base model (e.g., LLaMA) and attach multiple specialized LoRA adapters for different use cases. Example application areas include:

• CSRD (Corporate Sustainability Reporting Directive): Extraction and processing of sustainability-related data
• SRM (Supplier Relationship Management): Analysis of supplier and contract information
• Data Extraction: Automatic extraction of structured information from unstructured documents in an ESG context
• Synthetic Data Generation: automatic generation of training data for custom models

Objectives:
The goal is to design an efficient and scalable architecture that enables several LoRA modules to be applied on top of one shared base model, without linear increases in memory consumption. This should allow the deployment of multiple specialized models without redundant training or storage costs.

Methodological Focus:

• Application of modern machine learning techniques in the field of transfer learning and parameter-efficient fine-tuning
• Implementation and comparison of LoRA with alternative adapter-based methods
• Investigation of memory–performance trade-offs when using multiple LoRA modules simultaneously
• Development and evaluation based on realistic datasets

Requirements for Students:

• Solid background in machine learning and deep learning frameworks (e.g., PyTorch, HuggingFace Transformers)
• Understanding of LLM architectures and transfer learning methods
• Interest in bridging the gap between research and industrial application

This project offers students the opportunity to work on a cutting-edge research topic in LLM efficiency while producing results that are directly relevant for real-world industrial scenarios.