alinashrestha/patentsbert-m4-final

PatentSBERTa M4 Final — Y02 Green Patent Classifier

Fine-tuned version of AI-Growth-Lab/PatentSBERTa for binary classification of patent claims as GREEN (Y02) or NOT GREEN.

Developed as the final model in a 4-assignment pipeline for the course Applied Deep Learning and Artificial Intelligence at Aalborg University.

---

Model Performance

Dataset	Accuracy	Precision	Recall	F1
Eval Silver (10,000 claims)	0.8048	0.8025	0.8086	0.8055
Gold 100 (human verified)	0.5700	0.4310	0.7143	0.5376

---

Training Data

• 30,000 silver-labelled patent claims (Y02 CPC codes)
• 100 gold-labelled claims (human verified via HITL)
• Total: 30,100 training examples — perfectly balanced 50/50

---

Pipeline Overview

This model is the final step in a 4-stage pipeline:

1. QLoRA Fine-Tuning — Mistral-7B fine-tuned as Y02 domain expert Judge
2. 3-Agent MAS — Advocate / Skeptic / Judge debate on 100 uncertain claims
3. HITL Review — Human reviewed only 11 flagged uncertain claims
4. PatentSBERTa Fine-Tuning — Final classifier trained on 30,100 examples

---

Training Config

Parameter	Value
Base model	AI-Growth-Lab/PatentSBERTa
Epochs	1
Learning rate	2e-5
Batch size	16 train / 32 eval
Max sequence length	256
Hardware	NVIDIA L4 GPU
Training time	~6 minutes

---

Files in this Repository

File	Description
model.safetensors	Final model weights
gold_100_dataset.csv	100 gold labelled claims
train_silver_gold_30100.csv	Full training dataset
mas_100_results.csv	MAS agent outputs
hitl_100_gold.csv	Human reviewed labels
Train_qlora_mistral_v2.PY	QLoRA training script
mas_crewai_fewshot_v2.PY	MAS pipeline script
train_patentsbert_final.py	PatentSBERTa training script
m4_final_assignmnet.ipynb	Full assignment notebook

---

Citation

@misc{shrestha2026greenpatent,
  author = {Alina Shrestha},
  title  = {Green Patent Detection: Advanced Agentic Workflow with QLoRA},
  year   = {2026},
  note   = {AAU Applied Deep Learning and Artificial Intelligence}
}