Lightweight Fine-Tuning for Document Automation in the Cloud
By Gabriele Monti – MSc Data Science, Birkbeck, University of London
What This Is
This project explores how to fine-tune large language models (LLMs) efficiently for document manipulation tasks in cloud environments. The focus is on building resume parsers and PDF-to-JSON converters using lightweight methods that reduce training time and hardware requirements.
Key Use Cases
- Resume to structured JSON
- Invoice and form parsing
- Automatic document summarization
- PDF data extraction
- AI-powered document classification
Our Approach
We use parameter-efficient fine-tuning (PEFT) methods such as:
- LoRA (Low-Rank Adaptation)
- Adapter modules
- Prefix tuning
These approaches allow us to achieve high accuracy while remaining cost-effective and cloud-deployable.
What We Built
- Resume-to-JSON parser using Google Gemma + LoRA
- Full pipeline: data preparation, training, evaluation, deployment
- Lightweight deployment on Hugging Face Spaces and Docker
- Evaluation with WeightWatcher for layer quality and compression metrics
- API and Web Interface for real-time testing
Technologies Used
- Python + PyTorch
- Hugging Face Transformers & PEFT
- Google Colab, Docker, Hugging Face Spaces
- REST API (FastAPI) and frontend integration
Try It or See the Code
The full codebase is open-source:
GitHub Repository:
https://github.com/Birkbeck/msc-projects-2023-4-Gabriele_Monti_PEFT
Work With Us
Interested in deploying AI for document processing in your business?
Get in touch for a tailored solution.