UrbanGPT is a spatio-temporal large language model designed for urban computing tasks, targeting researchers and practitioners in AI and urban planning. It aims to enhance generalization capabilities in data-scarce environments by integrating spatio-temporal encoding with instruction tuning.

How It Works

UrbanGPT integrates a spatio-temporal dependency encoder (TCN-based) with an instruction-tuning paradigm. This approach allows LLMs to understand complex inter-dependencies across time and space, enabling more accurate predictions, particularly in zero-shot scenarios, by leveraging pre-trained spatio-temporal data.

Quick Start & Requirements

• Installation: Clone the repository and install dependencies using pip install -r requirements.txt. Key dependencies include PyTorch (2.0.1+cu117), fschat, torch_geometric, deepspeed, ray, einops, wandb, and flash-attn==2.3.5.
• Prerequisites: CUDA 11.7, Python 3.9.13.
• Data/Models: Requires downloading Vicuna v1.5 or v1.5-16k checkpoints and spatio-temporal training data (NYC taxi, bike, crime data).
• Resources: Training involves significant computational resources, indicated by multi-GPU usage (nnodes=1 --nproc_per_node=8).
• Links: Huggingface Models, Huggingface Datasets, Paper.

Highlighted Details

• Accepted to KDD'2024.
• Provides pre-trained models and datasets on Huggingface.
• Supports instruction tuning and evaluation for spatio-temporal tasks.
• Includes code for instruction generation for zero-shot and supervised prediction.

Maintenance & Community

The project is associated with the Data Intelligence Lab at the University of Hong Kong and South China University of Technology. Further community engagement details are not explicitly provided in the README.

Licensing & Compatibility

The repository does not explicitly state a license. The project acknowledges inspirations from Vicuna, GraphGPT, NExT-GPT, gradio, and Baize.

Limitations & Caveats

The README mentions potential version compatibility issues with flash-attn and transformers. The project is presented as a research artifact with a "TODO" list including releasing baseline codes.