This repository is part of my master's thesis on Entity Matching (EM). In my project, a parameter-efficient approach - Low-Rank Adaptation (LoRA) - is used to adapt pre-trained transformers for EM tasks. By updating only a small subset of parameters, this method achieves strong performance on unlabeled target datasets by aligning feature distributions between labeled source and unlabeled target datasets.

• 🔍 Problem: Entity Matching (EM) requires labeled training data, which is expensive to obtain. This project explores domain adaptation to transfer knowledge from labeled source datasets to unlabeled target datasets.
• 💡 Solution: Uses LoRA (Low-Rank Adaptation) for efficient fine-tuning and MK-MMD (Multiple Kernel Maximum Mean Discrepancy) for domain alignment.
• ✨ Key Features:
  • ✅ Efficient fine-tuning with LoRA (only 1% additional parameters).
  • ✅ Domain adaptation to align source and target feature distributions.
  • ✅ Task learning and domain adaptation trade-off, controlled by a dynamic α function.

│── config/                # YAML config files for training & hyperparameter tuning
│── dataset/               # Data processing and dataloaders
│── src/
│   ├── models/            # LoRA and Fine-Tuning models
│   ├── divergences/       # Domain adaptation losses (MK-MMD, Gaussian Kernels)
│   ├── training/          # Training scripts
│   ├── evaluation/        # Evaluation metrics and scripts
│── scripts/
│   ├── train.sh           # Training script
│   ├── run_experiments.sh # Batch execution for different datasets
│── Dockerfile             # Docker setup for containerized execution
│── requirements.txt       # Required Python dependencies
│── setup.sh               # Setup script for dependencies (CUDA support included)
│── README.md              # Project documentation

The script:

1. Loops through source and target dataset experimented in the paper, using 3 random seeds
2. Use the --hparam_tuning flag to enable hyperparameter tuning.

This project is containerized using Docker for consistency and reproducibility.
To set up and run the model inside a Docker container:

docker build -t thesis-peft .

docker run --gpus all --rm -v $(pwd):/app thesis-peft bash train.sh --src <source_dataset> --tgt <target_dataset> --model lora

docker run --gpus all --rm -v $(pwd):/app thesis-peft bash train.sh --src <source_dataset> --tgt <target_dataset> --model lora --hparam-tuning