sandner/n-body-trajectories-for-vrlora

N-Body Trajectories for Visual Reasoning Research

Dataset Description

This dataset contains a collection of simulated trajectories for a 3-body gravitational system. It was generated to train and evaluate "Visual Reasoning LoRA" (VR-LoRA) models as described in the paper: "Visual Reasoning Transfer: Leveraging Pretrained Visual Models for Physical and Temporal Prediction".

The core idea of the research is to encode these physical state vectors into "spatial field images" and train a vision model to predict the temporal evolution of the system.

• Research Project Repository: https://github.com/sandner-art/SC-Visual-Reasoning
• Trained Model (LoRA): huggingface.co/sandner/vr-lora-physics-sd15

How to Use this Dataset

The dataset is provided as compressed NumPy .npz files. It can be loaded directly with NumPy.

import numpy as np
from huggingface_hub import hf_hub_download

# Download the training data
file_path = hf_hub_download(
    repo_id="sandner/n-body-trajectories-for-vrlora", # Replace with your repo name
    filename="nbody_3_train_10000.npz",
    repo_type="dataset"
)

# Load the trajectories
data = np.load(file_path)
trajectories = data['trajectories']

# The shape is (num_samples, num_timesteps, num_particles, state_dim)
# e.g., (10000, 50, 3, 4)
print(trajectories.shape)

# Each state is a vector of [x, y, vx, vy]
first_trajectory_first_step = trajectories
print(first_trajectory_first_step)

Dataset Structure

The repository contains two primary files:

• nbody_3_train_10000.npz: The main training set with 10,000 trajectories.
• nbody_3_test_500.npz: A smaller test set with 500 trajectories for evaluation.

Each trajectory consists of 50 timesteps for a 3-particle system. Each particle's state at each timestep is represented by a 4-dimensional vector: [position_x, position_y, velocity_x, velocity_y].

Dataset Creation

The data was generated using a custom Python script (generate_dataset.py in the main GitHub repo) that utilizes scipy.integrate.solve_ivp with a Runge-Kutta (RK45) method.

Key simulation parameters:

• Gravitational Constant (G): 1.0
• Time Span: 0.0 to 2.0
• Softening Factor: 1e-3 (to prevent numerical singularities)
• Initial Conditions: Particle positions, velocities, and masses were sampled from uniform random distributions.

Citing this Dataset

If you use this dataset in your research, please cite our paper (BibTeX entry coming soon) and link back to this repository.