Parameter-Efficient and Quantized LLM Fine-Tuning for Predictive Process Monitoring

Under review at Information Systems.

---

This repository trains next-event prediction models on public BPI process event logs. It predicts the next activity and resource, plus the remaining time, using either an RNN/LSTM/GRU/Transformer or a pretrained Hugging Face language model with optional LoRA and quantization.

Reproduce

Python 3.12 and a CUDA-capable GPU are recommended.

uv venv .venv --python 3.12
source .venv/bin/activate
uv pip install -r requirements.txt

# Vanilla RNN experiment
python next_event_prediction.py configs/vanilla.yaml

# LoRA fine-tuning experiment
python next_event_prediction.py configs/fine_tuning.yaml

The event logs are downloaded automatically and cached in data/. Configuration values can be overridden from the command line, for example:

python next_event_prediction.py configs/vanilla.yaml training.epochs=10 model.backbone=transformer

Metrics are printed after every epoch. Set training.persist_best=true to save the best checkpoint under outputs/persisted_models/, or training.use_wandb=true to log runs to Weights & Biases.