The official implementation of ILRec. The implementation of ILRec is based on PyTorch and RecBole.

• We present ILRec, a novel LLM-based recommendation frame-work that extracts fine-grained self-hard negative signals from intermediate layers for preference optimization.
• We propose the cross-layer preference fine-tuning to penalize negative tokens based on the self-hard signals during LLM fine-tuning. Moreover, we introduce collaborative reward regularization to prevent over-penalty and integrate collaborative information.

1. Pre-processing

python data_processing.py

2. Train the model with ILRec framework LC-Rec:

bash ./train/ilrec-lcrec.sh

BIGRec:

bash ./train/ilrec-bigrec.sh