EminaGracanin/patentsberta-green-hitl

PatentSBERTa Green Classification (HITL Fine-Tuned)

This model is a fine-tuned version of AI-Growth-Lab/PatentSBERTa for binary classification of green patents.

Task

Binary classification:

• —0 = Not green
• —1 = Green (aligned with CPC Y02*)

Training Data

• —50,000 patent claims (balanced: 25k green / 25k non-green)
• —Silver labels derived from CPC Y02 codes
• —100 high-uncertainty examples manually reviewed via Human-in-the-Loop (HITL)
• —Gold labels override silver labels for those 100 examples

Methodology

1. 1.Baseline model using frozen embeddings + Logistic Regression
2. 2.Uncertainty sampling: u = 1 − 2 · |p − 0.5|
3. 3.Selection of 100 most uncertain claims
4. 4.LLM suggestion + Human review (HITL)
5. 5.Fine-tuning PatentSBERTa for 1 epoch (lr=2e-5, max_length=256)

Performance

Eval (silver split, 5k examples)

• —Accuracy: 0.806
• —Precision: 0.802
• —Recall: 0.820
• —F1: 0.811

Gold 100 (HITL se

Video link: https://aaudk-my.sharepoint.com/:v:/g/personal/xm37whstudentaau_dk/IQDKHMAiX07vSbSmU2pemhLFAbqbEx4kpKUsH4xu9LE55uQ?e=n5Xegc&nav=eyJyZWZlcnJhbEluZm8iOnsicmVmZXJyYWxBcHAiOiJTdHJlYW1XZWJBcHAiLCJyZWZlcnJhbFZpZXciOiJTaGFyZURpYWxvZy1MaW5rIiwicmVmZXJyYWxBcHBQbGF0Zm9ybSI6IldlYiIsInJlZmVycmFsTW9kZSI6InZpZXcifX0%3D