README.md

---
tags:
- patents
- climate
- green-technology
- text-classification
- patent-classification
- human-in-the-loop
- multi-agent
- patentsberta
language:
- en
pipeline_tag: text-classification
library_name: transformers
---

# Green Patent Detection: Multi-Agent HITL + PatentSBERTa

This repository contains an advanced green patent detection workflow built for **binary classification of patent claims** into:

- **1 = Green / climate mitigation related**
- **0 = Non-green**

The project extends a baseline PatentSBERTa workflow by adding a **Human-in-the-Loop (HITL)** review stage and a **multi-agent debate system** before final fine-tuning.

## Project overview

The goal of this project is to improve green patent detection by combining:

1. **High-risk sample selection** from uncertainty sampling
2. **Multi-agent LLM review** of difficult claims
3. **Human verification** of the AI suggestions
4. **Final fine-tuning of PatentSBERTa** using silver labels + gold HITL labels

This workflow was designed to test whether a more advanced labeling pipeline produces stronger training data than a simple single-LLM labeling approach.

## Base model

The final classifier is built from:

- **Base encoder:** `AI-Growth-Lab/PatentSBERTa`
- **Task:** Binary text classification
- **Domain:** Patent claim classification for climate mitigation / green technology

## Data used in the notebook

The notebook uses the following files:

- `patents_50k_green.parquet`
- `train_meta.csv`
- `y_train.npy`
- `eval_silver.parquet`
- `hitl_green_100.csv`
- `hitl_review_progress_with_llm.csv`
- `hitl_green_gold.csv`
- `hitl_three_agents.csv`

## Methodology

### 1. High-risk claim selection

A set of **100 high-risk patent claims** was selected from earlier uncertainty sampling outputs. These were the most difficult / ambiguous examples for the model.

### 2. Multi-agent debate system

Three agents were created using `CrewAI` and an Ollama-hosted model (`qwen2.5:3b-instruct`):

- **Advocate Agent** – argues why the claim should be classified as green under Y02 climate mitigation logic
- **Skeptic Agent** – argues why the claim may not qualify and checks for weak evidence or greenwashing
- **Judge Agent** – reviews both sides and returns a structured final output with:
  - predicted label
  - confidence
  - rationale

This produces an AI suggestion for each difficult claim.

### 3. Human-in-the-Loop review

The AI-generated suggestion was then manually reviewed.

The final human label was stored as:

- `is_green_human`

These human-reviewed labels form the **gold dataset** for the difficult claims.

### 4. Gold-enhanced training

The final training set combines:

- **Silver labels** from the earlier training data
- **100 gold human-reviewed claims** from the multi-agent workflow

This combined dataset was used to fine-tune PatentSBERTa.

## Training configuration

The notebook fine-tunes the model with the following setup:

- **Model:** `AI-Growth-Lab/PatentSBERTa`
- **Max sequence length:** `256`
- **Epochs:** `1`
- **Learning rate:** `2e-5`
- **Train batch size:** `8`
- **Eval batch size:** `8`
- **Weight decay:** `0.01`
- **Framework:** Hugging Face Transformers Trainer

## Dataset splits used during fine-tuning

From the notebook:

- **Training data:** silver training set + gold HITL labels
- **Evaluation data:** `eval_silver`
- **Additional check:** `gold_100`

The notebook text states that the final training dataset contains **35,200 claims**.

## Human vs AI agreement

According to the notebook:

- **Simple LLM from Assignment 2:** `94%` agreement with human labels
- **Agentic system from Assignment 3:** `87%` agreement with human labels

This suggests that the multi-agent system used stricter reasoning criteria, which created more disagreement on borderline cases.

## Repository contents

Depending on what you upload, this repository may include:

- the processed HITL dataset
- the final trained model
- tokenizer files
- training notebook
- prediction / rationale outputs for the 100 reviewed claims

## Expected columns in the HITL dataset

The notebook shows or creates columns such as:

- `id`
- `text`
- `p_green`
- `u`
- `llm_green_suggested`
- `llm_confidence`
- `llm_rationale`
- `is_green_human`

## Example use

```python
from transformers import AutoTokenizer, AutoModelForSequenceClassification
import torch

model_name = "YOUR_HF_REPO_NAME"
tokenizer = AutoTokenizer.from_pretrained(model_name)
model = AutoModelForSequenceClassification.from_pretrained(model_name)

text = "A company develops a carbon capture system that reduces CO2 emissions from cement factories."
inputs = tokenizer(text, return_tensors="pt", truncation=True, padding=True, max_length=256)

with torch.no_grad():
    logits = model(**inputs).logits
    pred = torch.argmax(logits, dim=-1).item()

print("Predicted label:", pred)
```

## Intended use

This project is intended for:

- research and coursework on green patent detection
- experimentation with HITL labeling pipelines
- comparison of simple vs advanced AI-assisted annotation workflows
- climate-tech related document classification

## Limitations

- The gold set is relatively small (**100 reviewed claims**)
- The multi-agent workflow depends on LLM reasoning quality
- Agreement with humans does not automatically guarantee better downstream model performance
- Final performance metrics should be reported from the actual training run in this repository

## Notes

This README was prepared from the notebook workflow and code structure. If you are uploading the **model repo**, add the final evaluation metrics from your training output. If you are uploading the **dataset repo**, you can keep the methodology sections and remove the model inference example if not needed.

## Citation

If you use this work, please cite the repository and the base model:

- `AI-Growth-Lab/PatentSBERTa`

You may also describe the workflow as:

> Multi-Agent Human-in-the-Loop green patent detection using PatentSBERTa with gold-enhanced fine-tuning.