alexchrander/patent-sberta-green-finetuned

Patent Green Technology Classifier (PatentSBERTa Fine-tuned)

A binary text classifier for detecting green/sustainable technology patent claims, built on top of AI-Growth-Lab/PatentSBERTa.

Model Description

This model was fine-tuned on a balanced dataset of 35,100 patent claims with gold-enhanced labels derived from a Human-in-the-Loop (HITL) workflow. It classifies patent claims as either green technology (1) or not green technology (0).

Training Data

• —Base dataset: AI-Growth-Lab/patents_claims_1.5m_traim_test
• —Silver labels: Derived from CPC Y02* codes (25,000 green + 25,000 not green)
• —Gold labels: 100 examples labeled via LLM → Human HITL workflow
• —Dataset: alexchrander/patents-green-gold-dataset

Training procedure

Active Learning + HITL workflow:

1. 1.Trained a frozen PatentSBERTa baseline using Logistic Regression
2. 2.Applied uncertainty sampling to identify the 100 most uncertain examples
3. 3.Used Mistral-7B-Instruct-v0.2 to suggest labels with rationale
4. 4.Human reviewer assigned final gold labels, overriding the LLM in 6/100 cases
5. 5.Fine-tuned PatentSBERTa on the gold-enhanced dataset

Hyperparameters:

• —maxseqlength: 256
• —epochs: 1
• —learning_rate: 2e-5
• —batch_size: 16

Results

HITL Override Examples

The human reviewer overrode the LLM suggestion in 6 out of 100 cases, all from not green → green:

1. 1.A phosphate detection method in soil and groundwater — labeled green as it relates to monitoring agricultural and water contamination
2. 2.A substrate coating method reducing film thickness — labeled green as material reduction can be considered a sustainable practice
3. 3.A surfactant removal method using electrochemical oxidation — labeled green as surfactant removal is associated with clean chemistry

Video

https://panopto.aau.dk/Panopto/Pages/Viewer.aspx?id=a519a0b3-17e2-44a5-b956-b3f90160c1c5