PatentSBERTa_finetuned_green

# Green Patent Dataset
## Overview
This dataset contains patent claim text labeled for green technology classification.

## Dataset Composition
Total size: 50,000 patents
Columns:
- doc_id
- text
- is_green_silver
- is_green_human (100 gold labels)
- is_green_final (gold override if available)
- split (train_silver / eval_silver)

Gold samples: 100 manually reviewed items.

---
## Label Definition
1 = Green Technology
0 = Non Green Technology

Green technology includes:
- Renewable energy
- Emission reduction
- Sustainability innovation
- Climate mitigation

---
## Evaluation Results

### Eval Silver
- F1: 0.8010

### Gold 100 (Human-labeled subset)
- F1: 0.5977

The drop in performance from Silver to Gold indicates that the human-corrected labels are more strict and realistic, highlighting the importance of Human-in-the-Loop refinement.
---

## Human-in-the-Loop (HITL) Analysis

### Override Statistics
Out of 100 reviewed samples:
- Total overrides: **21**
- Override rate: **21.0%**
This means that in 21% of cases, the human reviewer disagreed with the LLM's suggestion and corrected the label.
This highlights the importance of Human-in-the-Loop refinement, especially for ambiguous or borderline patent claims.

---
### Example Overrides
Below are three examples where the human label differs from the LLM suggestion:

---
#### Example 1
- **LLM Suggestion:** 1 (Green)
- **Human Label:** 0 (Not Green)
**Claim:**
"1. A standing pouch comprising: a first pack which comprises a first internal part and a second internal part; an accommodating space formed between the first internal part and the second internal part and configured to received and contain contents therein; a first external part which is disposed to face the first internal part to form a first temperature control space between the first external part and the first internal part; and a first temperature control member which is disposed in the first temperature control space, a first upper sealed part configured to seal an upper end of the first internal part and an upper end of the first external part to each other to close the first temperature control space; and a bottom external part sealing and coupled with a lower part of the first internal part and a lower part of the second internal part so that the lower part of the first internal part is spaced from the lower part of the second internal part, wherein the first upper sealed part is configured to be opened to allow an external reactant liquid to enter into the first temperature control space and react exothermically or endothermic ally with the first temperature control member disposes in the first temperature control space, the exothermic or endothermic reaction within the first temperature space configured to increase or decrease a temperature of contents within the accommodating space, wherein a first surface of the bottom external part is coupled with a lower internal surface of the first internal part and a lower internal surface of the second internal part respectively, so that a plane connecting between a lower end of the first internal part and a lower end of the second internal part is at least partially spaced apart from a second surface of the bottom external part, wherein the bottom external part is a discrete component separate from the first internal part and the second internal part, and coupled with a lower part of the first internal part and a lower part of the second internal part so as to be convex from a bottom surface on which the standing pouch is placed, wherein the bottom external part in combination with the first internal part and the second internal part define the accommodating space configured to receive and contain the contents to be heated or cooled."
**Reason:**
This claim describes a standing pouch designed to control the temperature of contents within an accommodating space through an exothermic or endothermic reaction. The key phrase "exothermic or endothermic reaction within the first temperature space configured to increase or decrease a temperature of contents" directly relates to temperature control, which can be used to preserve temperature-sensitive goods, aligning with sustainability goals. Furthermore, the pouch's design suggests a method for managing temperature, potentially reducing energy consumption associated with heating or cooling products.

---
#### Example 2
- **LLM Suggestion:** 1 (Green)
- **Human Label:** 0 (Not Green)
**Claim (shortened):**
"1. A method of detecting differences in density, comprising: providing a fluid medium with paramagnetic or superparamagnetic properties; introducing a diamagnetic material having a first density into the fluid medium to form a suspension of diamagnetic material in the fluid medium; in said suspension, exposing the diamagnetic material to a density modifying agent to form a modified diamagnetic material having a second density; and applying a magnetic field to the suspension, said suspension comprising at least the modified diamagnetic material, wherein the unmodified diamagnetic material and the modified diamagnetic material move to different positions in the fluid medium to occupy different positions in the magnetic field and the different positions correlate to a difference in density."
**Reason:**
This claim describes a method for detecting density differences using magnetic properties of materials in a fluid medium. The use of “paramagnetic or superparamagnetic properties” and “magnetic field” suggests a technology related to environmental monitoring or analysis, potentially for assessing water quality or contaminant detection – both fall under environmental protection. The manipulation of density through a “density modifying agent” further indicates a focus on characterizing and potentially mitigating environmental factors.

---
#### Example 3
- **LLM Suggestion:** 1 (Green)
- **Human Label:** 0 (Not Green)
**Claim (shortened):**
"1. A welding wire feeding apparatus, comprising: a control device which positions a welding gun a set distance above a work piece to be welded; a wire feeding device which feeds a wire to said welding gun; a power source which provides a current and a voltage to said wire while said wire feeding device is feeding said wire; and a short circuit detection circuit coupled to said wire feeding device, wherein when said short circuit detection circuit detects a short circuit in at least one of said wire and said welding gun, said wire feeding device stops feeding said wire based on said detection of said short circuit, and wherein said welding gun comprises a contact tip and said short circuit detection circuit causes said wire feeding device to stop said wire such that a stick out distance of said wire with respect to said contact tip is said set distance, said short circuit detection circuit including a power supply to apply a voltage to at least one of said wire and said contact tip and a voltage monitoring system for monitoring the voltage in the at least one of said wire and said contact tip and comparing the monitored voltage to a reference voltage."
**Reason:**
This claim describes a welding wire feeding apparatus with a short circuit detection system. The system’s purpose is to prevent damage to the welding gun and wire, which aligns with environmental protection by reducing waste and potentially improving the efficiency of welding processes. The inclusion of a voltage monitoring system suggests an attempt to optimize the welding process, a common goal within sustainable manufacturing.

---
## Intended Use
- Fine-tuning transformer models
- Studying label noise
- Human-in-the-loop learning
---
## Limitations
- Silver labels may contain noise
- Gold labels limited to 100 samples
---
## Disclaimer
This project was developed for academic purposes only. The classification results are intended for research and educational use, and should not be interpreted as legal advice or professional patent evaluation.
The Human-in-the-Loop (HITL) annotations were performed by students as part of a coursework assignment and do not represent expert legal judgment.
The model may contain biases and errors inherited from both automated labeling (silver labels) and LLM-assisted human review.
---
Video Link for Explaination: https://aaudk-my.sharepoint.com/:v:/g/personal/sm42zm_student_aau_dk/IQDAqRWONX2jTI9ftsBAgy8PAZvtvTZFe-Heffg_DEumxj0