Development of PEFT-Based Extraction Methodologies for Small-Scale LLMs

This thesis evaluates PEFT techniques, specifically LoRA and QLoRA, for adapting compact LLMs to extract structured JSON data from mechanical PDFs. By fine-tuning multiple architectures on a custom OCR and python extraction generated dataset, the study analyzes structural robustness, semantic accuracy, and hardware-induced computational trade-offs. The findings yield practical engineering guidelines for deploying efficient generative AI models in resource-bound industrial applications.