• 2024.7: We have resolved several bugs within our code. Below are the most recent results of LLaRA.

• 2024.5: We have updated the Steam dataset to a new version, in which we've addressed an issue that led to the repetition of certain data in the last interacted item of sequence.
• 🔥 2024.3: Our paper is accepted by SIGIR'24! Thank all Collaborators! 🎉🎉
• 🔥 2024.3: Our datasets and checkpoints are released on the huggingface.

1. Prepare the environment:

   git clone https://github.com/ljy0ustc/LLaRA.git
   cd LLaRA
   pip install -r requirements.txt

2. Prepare the pre-trained huggingface model of LLaMA2-7B (https://huggingface.co/meta-llama/Llama-2-7b-hf).

3. Download the data and checkpoints.

4. Prepare the data and checkpoints:

   Put the data to the dir path data/ref/ and the checkpoints to the dir path checkpoints/.

Train LLaRA with a single A100 GPU on MovieLens dataset:

sh train_movielens.sh

Train LLaRA with a single A100 GPU on Steam dataset:

sh train_steam.sh

Train LLaRA with a single A100 GPU on LastFM dataset:

sh train_lastfm.sh

Note that: set the llm_path argument with your own directory path of the Llama2 model.

Test LLaRA with a single A100 GPU on MovieLens dataset:

sh test_movielens.sh

Test LLaRA with a single A100 GPU on Steam dataset:

sh test_steam.sh

Test LLaRA with a single A100 GPU on LastFM dataset:

sh test_lastfm.sh