stratplans/savage-x3d-generation

Savage X3D Model Generation Dataset

Dataset Description

Purpose

This dataset enables training of Large Language Models (LLMs) to generate structured 3D models in X3D format from natural language descriptions. It contains 19,712 instruction-following examples derived from the Savage 3D Model Repository maintained by the Naval Postgraduate School.

Key Statistics

Metric	Value
Total Training Examples	19,712
Base X3D Models	1,232
Augmentation Factor	16x
Dataset Size	1.7 GB
Average Model Size	~90 KB
Total Downloaded Models	3,696
Processing Time	~30 minutes
Hardware Used	64-core, 5x RTX 4090

Quick Start

from datasets import load_dataset

# Load the dataset
dataset = load_dataset("stratplans/savage-x3d-generation")

# Example usage
example = dataset['train'][0]
print(f"Instruction: {example['instruction']}")
print(f"Input metadata: {example['input']}")
print(f"Output X3D: {example['output'][:500]}...")  # First 500 chars

Dataset Structure

Data Fields

Each training example contains four fields:

1. instruction (string): Natural language prompt requesting X3D generation

   • Example: "Create an X3D model of military ground vehicle with 4 shapes using Box, Cylinder geometries"

2. input (string): JSON-encoded metadata providing context

   {
     "category": "vehicle",
     "keywords": ["military", "hmmwv", "armored"],
     "complexity": "medium",
     "profile": "Immersive",
     "source_file": "HMMWV"
   }
   

3. output (string): Complete X3D XML model

   <?xml version="1.0" encoding="UTF-8"?>
   <!DOCTYPE X3D PUBLIC "ISO//Web3D//DTD X3D 3.0//EN" 
            "https://www.web3d.org/specifications/x3d-3.0.dtd">
   <X3D profile='Immersive' version='3.0'>
     <Scene>
       <Transform>
         <Shape>
           <Appearance>
             <Material diffuseColor='0.5 0.5 0.5'/>
           </Appearance>
           <Box size='2 1 4'/>
         </Shape>
       </Transform>
     </Scene>
   </X3D>
   

4. metadata (string): JSON-encoded additional information

   • Model complexity metrics (node count, shapes, materials)
   • Viewpoint information
   • Animation details
   • Original model ID

Dataset Categories

The dataset covers 9 major categories from the Savage repository:

Category	Description	Example Models
GroundVehicles	Military ground vehicles	HMMWV, M1A1, Jeep
AircraftFixedWing	Fixed-wing aircraft	F-16, F-18, AV8B Harrier
AircraftRotaryWing	Helicopters	Apache, BlackHawk, CH-46
ShipsMilitary	Naval vessels	Destroyers, Frigates, Carriers
ShipsCivilian	Civilian vessels	Tankers, Ferries, Tugboats
Buildings	Structures	Hangars, Houses, Stadiums
Sensors	Detection equipment	Radar, Sonar, Satellites
Weapons	Military ordnance	Missiles, Bombs
AmphibiousVehicles	Amphibious craft	LCAC, AAV

Dataset Creation Process

1. Data Collection (Parallel Download)

# Used 32 parallel threads for efficient downloading
with concurrent.futures.ThreadPoolExecutor(max_workers=32) as executor:
    # Downloaded 3,696 X3D files from savage.nps.edu
    # Crawled 1,439 index pages
    # Total download time: ~10 minutes

2. Data Processing Pipeline

# Parse X3D files and extract structured information
for x3d_file in x3d_files:
    metadata = extract_metadata(x3d_file)      # Title, creator, keywords
    scene_info = extract_scene_info(x3d_file)  # Shapes, materials, transforms
    complexity = calculate_complexity(x3d_file) # Node counts, animations
    description = generate_description(...)     # Natural language description

3. Data Augmentation Techniques

Applied 5 augmentation strategies to increase dataset diversity:

1. Color Variations: Modified HSV values of materials
2. Scale Transformations: Applied 0.5x to 2x scaling
3. Rotation Modifications: Added random rotations on axes
4. Viewpoint Adjustments: Modified camera positions
5. Lighting Variations: Adjusted light intensities and colors

4. Instruction Generation

Created diverse instruction templates:

templates = [
    "Create an X3D model of {description}",
    "Generate a 3D scene showing {description}",
    "Build an X3D representation of {description}",
    "Design an X3D model featuring {description}",
    "Construct a 3D model with {description}"
]

Training Configuration Example

from transformers import AutoModelForCausalLM, TrainingArguments

model = AutoModelForCausalLM.from_pretrained("Qwen/Qwen2.5-Coder-7B")

training_args = TrainingArguments(
    output_dir="./x3d-generator",
    num_train_epochs=3,
    per_device_train_batch_size=4,
    gradient_accumulation_steps=2,
    learning_rate=2e-5,
    bf16=True,
    gradient_checkpointing=True,
    logging_steps=100,
    save_strategy="epoch",
    evaluation_strategy="epoch"
)

Use Cases

1. 3D Content Generation: Automatically generate 3D models from text descriptions
2. Simulation & Training: Create scenarios for military/defense simulations
3. Education: Teach 3D graphics and X3D standards
4. Research: Benchmark structured data generation in LLMs
5. Game Development: Rapid prototyping of 3D assets

Limitations

• Domain Specific: Primarily military/defense-oriented models
• Format Specific: Only X3D format (not OBJ, FBX, GLTF)
• No Textures: Models don't include texture image files
• Geometric Focus: Limited artistic/organic shapes
• X3D Version: Mostly X3D 3.0/3.3 (not latest 4.0)

Citation

@dataset{savage_x3d_generation_2024,
  title={Savage X3D Model Generation Dataset},
  author={Web3D Consortium and Naval Postgraduate School},
  year={2024},
  publisher={Hugging Face},
  journal={Hugging Face Datasets},
  howpublished={\url{https://huggingface.co/datasets/stratplans/savage-x3d-generation}}
}

Acknowledgments

• Naval Postgraduate School for maintaining the Savage repository
• Web3D Consortium for X3D standards and tools
• Don Brutzman and the Savage development team
• All original model creators and contributors

License

Apache 2.0 - The dataset is provided under Apache 2.0 license. Original Savage models are provided under their respective licenses (see Savage License).

Links

• Savage Repository
• X3D Specifications
• Web3D Consortium
• Dataset Generation Code

Dataset Samples

Simple Example (Box):

<Shape>
  <Appearance>
    <Material diffuseColor='1 0 0'/>
  </Appearance>
  <Box size='2 2 2'/>
</Shape>

Complex Example (Vehicle with multiple parts):

<Transform translation='0 0.5 0'>
  <Shape>
    <Appearance>
      <Material diffuseColor='0.3 0.3 0.3'/>
    </Appearance>
    <Box size='4 1 2'/>
  </Shape>
</Transform>
<Transform translation='1.5 0 0.8'>
  <Shape>
    <Appearance>
      <Material diffuseColor='0.1 0.1 0.1'/>
    </Appearance>
    <Cylinder radius='0.3' height='0.2'/>
  </Shape>
</Transform>
<!-- Additional vehicle components... -->

---