ZoeYou/patentbert-pytorch

PatentBERT - PyTorch

BERT model specialized for patent classification using the CPC (Cooperative Patent Classification) system. (PyTorch version of the original PatentBert model.)

📊 Specifications

• Output classes: 656 (CPC subclass labels)
• Classification system: CPC (Cooperative Patent Classification)
• Architecture: BERT-base (768 hidden, 12 layers, 12 attention heads)
• Vocabulary: 30,522 tokens
• Format: SafeTensors

🏷️ CPC Classes (Real Distribution)

The model predicts classes according to the authentic CPC system used in PatentBERT training:

Main Sections (Actual Counts)

• A (84 classes): Human Necessities - Agriculture, Food, Health, Sports
• B (171 classes): Performing Operations; Transporting - Manufacturing, Transport
• C (88 classes): Chemistry; Metallurgy - Chemical processes, Materials
• D (40 classes): Textiles; Paper - Fibers, Fabrics, Paper-making
• E (31 classes): Fixed Constructions - Building, Mining, Roads
• F (101 classes): Mechanical Engineering; Lightning; Heating; Weapons; Blasting
• G (81 classes): Physics - Optics, Acoustics, Computing, Measuring
• H (51 classes): Electricity - Electronics, Power generation, Communication
• Y (9 classes): General Tagging of New Technological Developments

Example of CPC Subclasses

• A01B: SOIL WORKING IN AGRICULTURE OR FORESTRY
• B25J: MANIPULATORS; CHAMBERS PROVIDED WITH MANIPULATION DEVICES
• C07D: HETEROCYCLIC COMPOUNDS
• G06F: ELECTRIC DIGITAL DATA PROCESSING
• H04L: TRANSMISSION OF DIGITAL INFORMATION

🚀 Usage

from transformers import BertForSequenceClassification, BertTokenizer
import json
import torch

# Load model and tokenizer
model = BertForSequenceClassification.from_pretrained('ZoeYou/patentbert-pytorch')
tokenizer = BertTokenizer.from_pretrained('ZoeYou/patentbert-pytorch')

# Inference example
text = "A method for producing synthetic materials with enhanced thermal properties..."
inputs = tokenizer(text, return_tensors="pt", max_length=512, truncation=True, padding=True)

with torch.no_grad():
    outputs = model(**inputs)
    predictions = outputs.logits.softmax(dim=-1)

# Get prediction
predicted_class_id = predictions.argmax().item()
confidence = predictions.max().item()

# Use model labels (CPC codes)
predicted_label = model.config.id2label[str(predicted_class_id)]

print(f"Predicted CPC class: {predicted_label} (ID: {predicted_class_id})")
print(f"Confidence: {confidence:.2%}")

📁 Included Files

• model.safetensors: Model weights (420 MB)
• config.json: Configuration with integrated CPC labels
• vocab.txt: Tokenizer vocabulary
• tokenizer_config.json: Tokenizer configuration
• labels.json: Complete CPC label mapping (656 authentic labels)
• README.md: This documentation

🔬 Performance

This model was trained on a large patent corpus to automatically classify documents according to the CPC system, using the exact same 656 CPC codes from the original PatentBERT training data.

📖 References

• Cooperative Patent Classification (CPC)
• Original PatentBERT Paper

📝 Citation

If you use this model, please cite the original PatentBERT work and mention this PyTorch conversion.

@article{patent_bert, 
  author = "Jieh-Sheng Lee and Jieh Hsiang",
  title = "{PatentBERT: Patent classification with fine-tuning a pre-trained BERT model}",
  journal = "World Patent Information",
  volume = "61",
  number = "101965",
  year = "2020",
}