LoRA_LLM_training

Green Patent Classification using LoRA Fine-Tuning

Overview

This project implements a parameter-efficient fine-tuning approach to classify patents as environmentally sustainable ("green") or not green. A pretrained DistilGPT2 model was adapted using Low-Rank Adaptation (LoRA) to create an LLM-based patent judge.

The model is trained to output a binary decision (0 or 1) based on patent text input.

---

Objective

The goal of this assignment is to:

1. Apply parameter-efficient fine-tuning (LoRA) to a pretrained language model.
2. Build an LLM-based classifier for green technology patents.
3. Generate machine-assisted supervision (MAS) labels.
4. Improve dataset quality by overriding silver labels with LLM-generated gold labels.

---

Base Model

• Base model: distilgpt2
• Total parameters: ~82M
• Trainable parameters (LoRA): ~147K (~0.18%)

Only LoRA adapter weights were trained. The base model remains frozen.

---

Methodology

1. Dataset

• Source: 50,000 patent documents
• Label type: Binary (is_green)
• Training split: Silver-labeled data
• Evaluation subset: Top 100 high-priority patents

2. Fine-Tuning Strategy

• Applied LoRA using the PEFT library
• Supervised fine-tuning (SFT)
• Instruction-style prompt formatting
• Binary JSON output format

3. LLM-Based Labeling

The trained model was used as a judge with the following prompt structure:

"You are a patent judge. Return ONLY JSON with key is_green (0 or 1)."

The model generated labels for 100 selected patents, creating a gold-standard subset.

---

Results

Gold label distribution (100 patents):

• Non-green (0): 53
• Green (1): 47

The generated gold labels were used to override corresponding silver labels in the full dataset, improving label quality.

---

Key Contributions

• Demonstrates parameter-efficient fine-tuning
• Builds a domain-specific LLM classifier
• Implements machine-assisted supervision
• Shows practical use of LoRA for real-world NLP tasks

---

Limitations

• Binary classification may oversimplify sustainability
• LLM outputs occasionally require fallback parsing
• Silver labels may still contain noise outside the gold subset

---

Future Improvements

• Use larger base models
• Apply 8-bit or 4-bit quantization for efficiency
• Add human validation for gold labels
• Evaluate performance using precision/recall metrics

---

Usage

To load the model with PEFT:

from transformers import AutoModelForCausalLM
from peft import PeftModel

base_model = AutoModelForCausalLM.from_pretrained("distilgpt2")
model = PeftModel.from_pretrained(base_model, "your-username/your-repo-name")

---

Author

Assignment submission for Green Patent Classification using LLMs and LoRA fine-tuning.

Link to video

https://aaudk-my.sharepoint.com/:v:/r/personal/ox26dw_student_aau_dk/Documents/Optagelser/Meeting%20in%20Group%20AAE-20260302_121642-Meeting%20Recording.mp4?csf=1&web=1&e=EX49U0&nav=eyJyZWZlcnJhbEluZm8iOnsicmVmZXJyYWxBcHAiOiJTdHJlYW1XZWJBcHAiLCJyZWZlcnJhbFZpZXciOiJTaGFyZURpYWxvZy1MaW5rIiwicmVmZXJyYWxBcHBQbGF0Zm9ybSI6IldlYiIsInJlZmVycmFsTW9kZSI6InZpZXcifX0%3D