prithivMLmods/Text-to-Code-3D-Samples

Text-to-Code-3D-Samples

Text-to-Code-3D-Samples is a medium-scale multimodal 3D dataset containing ~59,999 paired entries of natural language design specifications, executable OpenSCAD (Open-source Solid 3D CAD Modeller) programmatic source scripts, and 3D visual render representations. Converted and curated into Apache Parquet format from publicly accessible CAD model repositories, open-source 3D asset libraries, and synthetic parametric definitions, this dataset is designed for training, supervised fine-tuning (SFT), and evaluating Code-LLMs and Vision-Language Models (VLMs) on parametric CAD synthesis (Text-to-CAD, Image-to-CAD, and CAD code generation).

This is an experimental dataset only and may contain artifacts.

• Total Samples: ~59,999 rows
• Format: Apache Parquet (Prompt, Code, Image)
• Modalities: Text, Code, Image
• Split: Train

Dataset Structure & Schema

Feature Fields

Field	Type	Description
Prompt	string	Technical description detailing dimensions, physical functional parts, and structural topology
Code	string	Executable programmatic OpenSCAD code script implementing constructive solid geometry (CSG)
Image	Image	Rendered 2D isometric/orthographic preview of the generated 3D solid model

Data Instance Example

{
  "Prompt": "Design a parametric mounting bracket with four counter-sunk mounting holes along a rectangular base and a reinforced center rib.",
  "Code": "difference() { union() { cube([60, 40, 5], center=true); translate([0, 0, 15]) cube([10, 40, 25], center=true); } for(x=[-20, 20], y=[-12, 12]) translate([x, y, -3]) cylinder(h=8, r=2.5, $fn=30); }",
  "Image": "<PIL.PngImagePlugin.PngImageFile image mode="RGB" size="256x256">"
}

How to Use

Loading with datasets

from datasets import load_dataset

# Load the dataset
dataset = load_dataset("prithivMLmods/Text-to-Code-3D-Samples", split="train")

# Access a single record
sample = dataset[0]
prompt = sample["Prompt"]
code = sample["Code"]
rendered_image = sample["Image"]

print("Prompt:\n", prompt)
print("\nOpenSCAD Code:\n", code)

Instruction Fine-Tuning Format

def format_cad_instruction(example):
    instruction = (
        "You are an expert mechanical engineer and CAD programmer. Write compilable "
        "OpenSCAD code that builds the 3D geometry specified below.\n\n"
        f"Specification: {example['Prompt']}"
    )
    
    return {
        "prompt": instruction,
        "completion": example["Code"]
    }

formatted_dataset = dataset.map(format_cad_instruction)

Intended Uses

• Parametric CAD Generation: Fine-tuning Code-LLMs (e.g., CodeLlama, Qwen-Coder, DeepSeek-Coder) to translate natural language engineering requests into compilable OpenSCAD scripts.
• Multimodal Reverse-CAD: Training Vision-Language Models to reverse-engineer constructive solid geometry (CSG) code directly from multi-view or isometric renders.
• Rapid Prototyping & Additive Manufacturing: Providing structured references for automating custom 3D printing components, brackets, enclosures, and adapters.