rLLM

rLLM (relation LLM) focuses on LLM-based relational data mining, prioritizing: Accuracy, Efficiency, and Economy.

• Accuracy: MAE for regression; Micro-F1 and Macro-F1 for classification.
• Efficiency: Runtime, measured in seconds.
• Economy: Money, measured in dollars.

Environment Setup

Due to variations in the configurations of everyone's computer, achieving uniform setup is not feasible. Therefore, the following instructions address potential installation issues:

It is recommended to use a Linux system for experimentation, which also facilitates submission.

For Windows systems, installing WSL is advised. But you can also use your own system.

PyTorch Installation

• PCs with Nvidia GPUs can use the nvidia-smi command to check their CUDA support version.
• PCs without dedicated Nvidia GPUs should install the CPU version.
• PyTorch official website

llama-cpp-python and langchain

• Default installation method: CPU only (Windows/Linux/MacOS)

pip install llama-cpp-python

• If you want to use GPU, you need to first install CUDA and then install llama-cpp-python:

This allows specifying the n_gpu_layers parameter when instantiating the llama object, which determines how many layers of parameters are placed on the GPU to accelerate runtime.

# Instructions for installing GPU-enabled llama-cpp-python on Linux
# First, install the CUDA Toolkit. Tutorial: https://blog.csdn.net/qq_32033383/article/details/135015041. CUDNN installation is not necessary.

# Then use the following command
export LLAMA_CUBLAS=1
CMAKE_ARGS="-DLLAMA_CUBLAS=on" FORCE_CMAKE=1 pip install --upgrade --force-reinstall llama-cpp-python --no-cache-dir

For detailed instructions, refer to abetlen/llama-cpp-python: Python bindings for llama.cpp (github.com)

• We have used langchain, langchain-experimental and langchain-community packages. For detailed instructions, refer to Introduction | 🦜️🔗 Langchain

Download 4-bit quantized LLM models

• Download the 4-bit quantized LLM models directly from the SJTU cloud storage. Currently, due to its great performance, we choose 4-bit quantized gemma 2b model as our LLM.
• Download gemma-2b-it-q4_k_m.gguf from gemma-2b-it-q4_k_m.gguf

Choosing Embedding Models

• If you need to use the BERT model for sentence embedding, it is recommended to use sentence-transformers/all-MiniLM-L6-v2 · Hugging Face
• Downloads can be obtained from the SJTU cloud storage, or directly from Hugging Face.
• Use Sentence-Transformers or HuggingFace Transformers library to invoke the model.
• You can also use LLM to make sentence embedding.