WeChatPay-Pretraining/PANTHER ? reverse-engineered prompt

Build me a Python project for sequential user behavior modeling on credit card transaction data, based on the PANTHER idea.

I want a workflow where I can preprocess the raw transactions, turn them into behavior sequences, pretrain a model on those sequences, and then reuse the learned embeddings for a fraud detection task. Please include the training and inference scripts, plus a simple fraud detection training script that uses the pretrained representations.

Make it easy to point the preprocessing at a raw Kaggle dataset path, then run the full pipeline from preprocessing to pretraining to embedding extraction and fraud model training. Keep the setup reproducible, with clear config files for the main experiments and sensible defaults so I can run it without a lot of manual setup. If you need to look up current docs or best practices for any libraries, go ahead and do that.