README.md

metadata

license: apache-2.0
task_categories:
  - image-to-image
tags:
  - lance
  - pcb-routing
  - flux
size_categories:
  - n<1K

Zero-Obstacle High-Density PCB Routing Dataset (z01)

Image-to-image dataset for fine-tuning Flux 2 Klein 4B to route PCB traces.

Task

Given a connection-pairs image showing color-matched pin pairs on a 10x10 mm board, produce a routed image with traces connecting each pair.

Edit instruction:

Route the traces between the color matched pins, using red for the top layer and blue for the bottom layer. Add vias to keep traces of the same color from crossing.

Format

Lance dataset with the following columns:

Column	Type	Description
id	string	Sample identifier (e.g. sample-000001)
input_image	large_binary	Connection-pairs PNG
output_image	large_binary	Routed PNG
edit_instruction	string	The edit instruction

Splits

Split	Samples
train	71
test	18

Usage

import lance

# Load from HuggingFace
train = lance.dataset("hf://datasets/makeshifted/zero-obstacle-high-density-z01/data/train.lance")
test  = lance.dataset("hf://datasets/makeshifted/zero-obstacle-high-density-z01/data/test.lance")

# Read a sample
row = train.take([0]).to_pylist()[0]
print(row["id"], row["edit_instruction"])

# Save the images
with open("input.png", "wb") as f:
    f.write(row["input_image"])
with open("output.png", "wb") as f:
    f.write(row["output_image"])