CTB2001/patents-green-50k

Green Patent Claims — 50k Balanced Dataset

A balanced binary-classification dataset of 50,000 US patent first claims labelled as green technology (1) or not green (0), plus 100 human-verified gold labels from a targeted HITL review.

Created for the Applied Deep Learning (AAU, Spring 2025) exam assignment on active learning, multi-agent silver labelling, and human-in-the-loop verification.

Dataset Details

Property	Value
Total examples	50,000
Label balance	25,000 green / 25,000 not-green
Text field	First claim of each patent
Label source	Silver (CPC Y02 codes)
Date range	2013-01-01 — 2017-12-26
Median text length	900 characters
Format	Parquet

Splits

Split	Rows	Green	Not-green	Purpose
train_silver	25,000	12,500	12,500	Model training
pool_unlabeled	15,000	7,500	7,500	Active-learning pool (uncertainty sampling)
eval_silver	10,000	5,000	5,000	Held-out evaluation

The split assignment is stored in the split column.

Columns

Column	Type	Description
id	int	Patent document ID
date	date	Patent publication date
text	str	Full text of the patent's first claim
is_green_silver	int	Binary silver label: 1 = green, 0 = not green
split	str	Dataset split: train_silver, pool_unlabeled, or eval_silver
Y02A	int	CPC subclass flag — Adaptation to climate change
Y02B	int	CPC subclass flag — Climate-change mitigation (buildings)
Y02C	int	CPC subclass flag — Carbon capture & storage
Y02D	int	CPC subclass flag — ICT for reduced energy consumption
Y02E	int	CPC subclass flag — Clean energy generation/storage
Y02P	int	CPC subclass flag — Climate-change mitigation (production)
Y02T	int	CPC subclass flag — Transportation
Y02W	int	CPC subclass flag — Waste management

Green Subcategory Distribution (Y02)

Subcategory	Count	Description
Y02E	7,400	Clean energy generation & storage
Y02T	6,757	Transportation
Y02P	4,809	Production & processing
Y02D	3,884	ICT energy reduction
Y02B	2,938	Buildings
Y02A	2,093	Adaptation to climate change
Y02W	787	Waste management
Y02C	410	Carbon capture & storage

A patent may belong to multiple Y02 subcategories.

Gold Labels

The file final_gold_labels.csv contains 100 human-verified gold labels produced during Part D (HITL review) of the exam pipeline. These samples were selected from the uncertainty pool using a LangGraph Multi-Agent System (Advocate → Skeptic → Judge → Exception) and flagged for human review based on low confidence or agent disagreement.

Column	Description
doc_id	Patent document ID
text	Patent claim text
agent_judge_label	MAS Judge's predicted label
confidence	Judge confidence score
is_green_gold	Human-verified gold label
human_notes	Reviewer notes

Gold label distribution: 27 green / 73 not-green.

Usage

import pandas as pd

df = pd.read_parquet("hf://datasets/CTB2001/patents-green-50k/patents_50k_green.parquet")

train = df[df["split"] == "train_silver"]
pool  = df[df["split"] == "pool_unlabeled"]
eval_ = df[df["split"] == "eval_silver"]

Or with 🤗 Datasets:

from datasets import load_dataset
ds = load_dataset("CTB2001/patents-green-50k")

Associated Model

The fine-tuned classifier trained on this dataset: CTB2001/PatentSBERTa-green-classifier

Labelling Methodology

1. Silver labels are derived from CPC Y02 classification codes assigned by patent offices — a patent is labelled green (1) if it carries any Y02 subclass code
2. Gold labels were produced via a multi-stage pipeline: uncertainty sampling → QLoRA-powered Multi-Agent System deliberation → human-in-the-loop verification of the 100 most critical cases

Limitations

• Silver labels rely on CPC Y02 codes, which may contain noise (false positives/negatives from patent office classification)
• Only 100 gold labels were human-verified — insufficient for standalone gold-standard evaluation
• All patents are in English; other languages are not represented
• Date range covers 2013–2017 only

Citation

@misc{trost-bertelsen2025patents-green-50k,
  author       = {Trøst-Bertelsen, Christian},
  title        = {Green Patent Claims — 50k Balanced Dataset},
  year         = {2025},
  howpublished = {Hugging Face Dataset Hub},
  url          = {https://huggingface.co/datasets/CTB2001/patents-green-50k}
}

Author

Christian Trøst-Bertelsen — Aalborg University, Student ID 20224083 Course: Applied Deep Learning, 8th semester, Spring 2025