Towards principled evaluation and exploration of long-range capabilities in graph learning.
The task is Energy prediction in the Ising-model, which is central to simulations with non-trivial system behavior.
Long, non-trivial correlations at critical temperature where far nodes impact the enrgy prediction. Local-only predictions are limited compared to oracle predictor.
Chaning the size and sigma of the system directly impacts task difficulty, allowing systematic evaluation and control across complexity levels.
Get a continous feedback signal both during training and evaluation of your models that goes beyond "binary all or nothing".
Diverse collection of graph datasets for comprehensive evaluation
dataset = LRIM(name='lrim_16_0.6_10k')
Read the full LRIM Graph Benchmark paper with detailed methodology, experimental results, and long-range analysis.
View Paper →Download the complete benchmark datasets with all size variations and difficulty levels directly on HuggingFace or use the provided PyG Dataloader.
HuggingFace Datasets →Get example code, include the dataset in your own codebase with the provided dataloader or inspect model checkpoints.
GitHub Repository →Quick setup to start using the LRIM Graph Benchmark
# Clone the entire repository (includes training and examples)
git clone https://github.com/iJorl/lrim_graph_benchmark.git
cd lrim_graph_benchmark/example
./setup.sh
# Or just download the dataloader
wget https://raw.githubusercontent.com/iJorl/lrim_graph_benchmark/main/example-setup/lrim_loader.py# wget https://raw.githubusercontent.com/iJorl/lrim_graph_benchmark/main/example-setup/lrim_loader.py
# Requires PyTorch Geometric and HuggingFace
from lrim_loader import LRIM
from torch_geometric.loader import DataLoader
# Load dataset
dataset = LRIM(root='data/', name='lrim_16_0.6_10k')
splits = dataset.get_idx_split()
# Create data loaders
train_loader = DataLoader(dataset[splits['train']], shuffle=True)
val_loader = DataLoader(dataset[splits['val']])
test_loader = DataLoader(dataset[splits['test']])Loading example code...Dive deeper into interactive visualizations, live benchmarks, and detailed analysis tools.
Launch Explorer