**This repository includes several methods of Explainable Recommendations (ExpRec): **
| Model | Core | Url |
|---|---|---|
| CER | Transformer | https://github.com/JMRaczynski/CER |
| PETER | Transformer | https://github.com/lileipisces/PETER |
| PEPLER | GPT-2 | https://github.com/lileipisces/PEPLER |
| RecLLM | GPT-2 and LLaMa-2 | https://github.com/XiuzeZhou/ExpRec |
- python>=3.9
- numpy==1.26.4
- torch==2.2.0
- transformers==4.40.2
Git clone our repository, creating a python environment and coping it to .sh files.
Code Structure:
├──data: datasets, download from step 3
├── ClothingShoesAndJewelry
├── MoviesAndTV
├── TripAdvisor
├── Yelp
├──llm: pretrained LLMs, download from step 2
├── gpt2-small
├── gpt2-medium
├── gpt2-large
├── llama-2
├──logs: training logs and results.
├──models: save the final model.
├──outputs: generated texts by trained models.
├──shell: .sh for running
├── train_mf.sh
├── train_recllm-gpt.sh
├── train_recllm-llm.sh
├── train_peter.sh
├── train_pepler.sh
├── train_cer.sh
- gpt2-small: https://huggingface.co/openai-community/gpt2
- gpt2-medium: https://huggingface.co/openai-community/gpt2-medium
- gpt2-large: https://huggingface.co/openai-community/gpt2-large
- llama-2: https://huggingface.co/meta-llama/Llama-2-7b
Datasets to download
- ClothingShoesAndJewelry:
- MoviesAndTV: Amazon Movies & TV
- TripAdvisor: TripAdvisor Hong Kong
- Yelp: Yelp 2019
sh ./shell/train_cer.sh 0
sh ./shell/train_peter.sh 0
sh ./shell/train_pepler.sh 0
The training of RecLLM contains two stages:
1. Build user/item embeddings by MF
sh ./shell/train_mf.sh 0
2. LoRA tuning for all datasets
GPT-2:
sh ./shell/train_recllm-gpt.sh 0
LLaMa-2:
sh ./shell/train_recllm-llama.sh 0
Please cite our paper if they are helpful to your work!