README.md

Kaggle PII Data Detection Competition - 1st Place Solution Full Code

This repository contains the code and configurations for our winning solution in the PII Data Detection competition hosted by The Learning Agency Lab. Our team's approach and results are detailed in the competition discussion on Kaggle.

Inferencing

The inference process is detailed in the notebook available at Kaggle: PII 1st Place Solution. It outlines the procedure for using the trained models to predict and process test data.

Table of Contents

• Kaggle PII Data Detection Competition - 1st Place Solution Full Code
  • Inferencing
  • Table of Contents
  • Setup
    • Hardware
    • Software
    • Dependencies
    • Datasets
  • Training

Setup

Hardware

• Instance: Ubuntu 20.04.5 LTS (128 GB boot disk)
• CPU: Intel(R) Xeon(R) Silver 4216 @ 2.10GHz (7 vCPUs)
• GPU: 8 x NVIDIA A100 40GB

Software

• Python: 3.10.13
• CUDA: 12.1

Dependencies

Clone the repository and install the required Python packages:

git clone https://github.com/bogoconic1/pii-detection-1st-place.git
cd pii-detection-1st-place
pip install -r requirements.txt

Datasets

Ensure the Kaggle API is installed and set up. Use the following script to download the necessary datasets:

sh ./setup_datasets.sh

Note: The script creates a data folder in the parent directory and downloads external datasets there.

Training

Our solution involves five Deberta-v3-large models, incorporating different architectures for diversity and performance. Below are some variants and their training commands:

• Multi-Sample Dropout Custom Model: Improves training stability and performance.

  python train_multi_dropouts.py

• BiLSTM Layer Custom Model: Adds a BiLSTM layer to enhance feature extraction, includes specific initialization to prevent NaN loss issues.

  python train_bilstm.py

• Knowledge Distillation: Utilizes well-performing models as teachers to boost a student model's performance, leveraging disparate datasets. It requires a teacher model. We used the best of Multi-Sample dropout models. Note: it requires a teacher model to be distlled with. We used the best of multi-sample dropout models.

  python train_distil.py

• Experiment 073: Uses augmented data with name swaps.

  python train_exp073.py

• Experiment 076: Introduces a random addition of consequential names to training data.

  python train_exp076.py