BIGRec

This is the implementatino of our work A Bi-Step Grounding Paradigm for Large Language Models in Recommendation Systems

For item embedding, due to the quota of the git LFS, you can use the link with password 0g1g.

Results and Models

Recently, we have trained Qwen2-0.5B on five Amazon datasets, the model parameters are presented in the following table:

|Dataset|Link| |----------------|----------------|----------------|----------------| |CDs_and_Vinyl|link| |Video_Games|link| |Toys_and_Games|link| |Sports_and_Outdoors|link| |Books|link|

For more details on data processing methods, recommendation performance and additional training information, please refer to this repo.

Environment

pip install -r requirements.txt

Preprocess

Please follow the process.ipynb in each data directory.

Training on Single Domain

for seed in 0 1 2
do
    for lr in 1e-4
    do
        for dropout in 0.05    
        do
            for sample in 1024
            do
                echo "lr: $lr, dropout: $dropout , seed: $seed,"
                CUDA_VISIBLE_DEVICES=$1 python train.py \
                    --base_model YOUR_LLAMA_PATH/ \
                    --train_data_path "[\"./data/movie/train.json\"]"   \
                    --val_data_path "[\"./data/movie/valid_5000.json"]" \
                    --output_dir /model/movie/${seed}_${sample} \
                    --batch_size 128 \
                    --micro_batch_size 4 \
                    --num_epochs 50 \
                    --learning_rate $lr \
                    --cutoff_len 512 \
                    --lora_r 8 \
                    --lora_alpha 16\
                    --lora_dropout $dropout \
                    --lora_target_modules '[q_proj,v_proj]' \
                    --train_on_inputs \
                    --group_by_length \
                    --resume_from_checkpoint 'XXX' \
                    --seed $seed \
                    --sample $sample 
            done    
        done
    done
done

Training on Multi Domain

for seed in 0 1 2
do
    for lr in 1e-4
    do
        for dropout in 0.05    
        do
            for sample in 1024
            do
                echo "lr: $lr, dropout: $dropout , seed: $seed,"
                CUDA_VISIBLE_DEVICES=$1 python train.py \
                    --base_model YOUR_LLAMA_PATH/ \
--train_data_path "[\"./data/movie/train.json\", \"./data/game/train.json\"]"  \
                    --val_data_path "[\"./data/movie/valid_5000.json\", \"./data/game/valid_5000.json\"]"  \
                    --output_dir ./model/multi/${seed}_${sample} \
                    --batch_size 128 \
                    --micro_batch_size 4 \
                    --num_epochs 50 \
                    --learning_rate $lr \
                    --cutoff_len 512 \
                    --lora_r 8 \
                    --lora_alpha 16\
                    --lora_dropout $dropout \
                    --lora_target_modules '[q_proj,v_proj]' \
                    --train_on_inputs \
                    --group_by_length \
                    --resume_from_checkpoint 'XXX' \
                    --seed $seed \
                    --sample $sample 
            done    
        done
    done
done

Training on Multiple GPU Card

We provide our accelerate config in ./config/accelerate.yaml

accelerate config # Please set up your config
for seed in 0 1 2
do
    for lr in 1e-4
    do
        for dropout in 0.05    
        do
            for sample in 1024
            do
                echo "lr: $lr, dropout: $dropout , seed: $seed,"
                CUDA_VISIBLE_DEVICES=0,1,2,3 accelerate launch train.py \
                    --base_model YOUR_LLAMA_PATH/ \
--train_data_path "[\"./data/movie/train.json\", \"./data/game/train.json\"]"  \
                    --val_data_path "[\"./data/movie/valid_5000.json\", \"./data/game/valid_5000.json\"]"  \
                    --output_dir ./model/multi/${seed}_${sample} \
                    --batch_size 128 \
                    --micro_batch_size 4 \
                    --num_epochs 50 \
                    --learning_rate $lr \
                    --cutoff_len 512 \
                    --lora_r 8 \
                    --lora_alpha 16\
                    --lora_dropout $dropout \
                    --lora_target_modules '[q_proj,v_proj]' \
                    --train_on_inputs \
                    --group_by_length \
                    --resume_from_checkpoint 'XXX' \
                    --seed $seed \
                    --sample $sample 
            done    
        done
    done
done

Inference

#  Taking movie as an example
python inference.py \
    --base_model YOUR_LLAMA_PATH/ \
    --lora_weights YOUR_LORA_PATH \
    --test_data_path ./data/movie/test/test_5000.json \
    --result_json_data ./movie_result/movie.json

Evaluate

# Taking Game as an example
# Directly
python ./data/movie/evaluate.py --input_dir ./movie_result
# CI Augmented
python ./data/movie/adjust_ci.py --input_dir ./movie_result # Note that you need to have your own SASRec/DROS model (Specify the path in the code)

If you're using this code in your research or applications, please cite our paper using this BibTeX:

@article{bao2023bi,
  title={A bi-step grounding paradigm for large language models in recommendation systems},
  author={Bao, Keqin and Zhang, Jizhi and Wang, Wenjie and Zhang, Yang and Yang, Zhengyi and Luo, Yancheng and Chen, Chong and Feng, Fuli and Tian, Qi},
  journal={arXiv preprint arXiv:2308.08434},
  year={2023}
}