Pattern2Code

Benchmark for the ASE 2026 paper "Pattern Over Pixels: Measuring Pattern Completion Bias in Multimodal Code Generation."

Project page: https://pattern2code.github.io/ · Artifact (Zenodo): https://doi.org/10.5281/zenodo.19341952

Pattern2Code measures visual pattern-completion bias in screenshot-to-code generation: when a webpage contains a repeated UI pattern and exactly one element deviates from it, does a multimodal LLM write the code that matches the pixels, or the code that matches the pattern?

Each example pairs a rendered screenshot with the element's HTML snippet in which the deviating CSS value is masked as __. The model must recover the masked value from the image.

Key results (5 frontier MLLMs)

Family	Mean accuracy	Mean bias rate
Card patterns (width)	21.17%	69.78%
Text patterns (font-size)	7.89%	80.22%

Bias rate = fraction of predictions equal to the repeated baseline 100%.

Contents

• 720 perturbed instances = 30 webpages × 2 pattern families × 3 positions × 4 magnitudes
• 360 structural-card (width) + 360 text-style (font-size) examples
• 1,440 screenshots — every instance rendered under 2 conditions (standard_image, noise_image)
• Perturbation magnitudes: 80%, 90%, 110%, 120%; unperturbed baseline is 100%

Layout

data/train-*.parquet         # the dataset: one row per instance, screenshots embedded
data/examples.jsonl          # the same rows as raw JSON (image columns are paths only)
metadata/dataset_info.json   # summary statistics
metadata/validation_report.json
html/{cards,texts}/          # perturbed HTML
images/{cards,texts}/{standard,noise}/   # the screenshots as loose PNG files

The Parquet files carry the screenshots inline, so both image columns render directly in the dataset viewer and decode to PIL.Image with no extra setup. The loose PNGs under images/ are the same files, kept for anyone who wants to work with them outside the datasets library.

Row schema

Field	Description
standard_image	Image — the rendered screenshot
noise_image	Image — the same page with eight opaque rectangles overlaid
example_id	Unique instance id
source_id	Seed webpage id (Design2Code)
pattern_family	structural_card or text_style
pattern_name, pattern_description, element_description	The repeated pattern being probed
attribute_name	Masked CSS property (width or font_size)
attribute_token	The mask token (__)
position	Which element was perturbed: first, mid, last
modified_index	Index of the perturbed element within the pattern
target_value	Ground truth — the perturbed value (e.g. 120%)
bias_value	The pattern-consistent baseline (always 100%)
code_snippet_line	The masked element's HTML
code_snippet_context	Masked element plus its sibling elements (model input)
html_path	Relative path to the full perturbed HTML
standard_image_path / noise_image_path	Relative paths to the two screenshot files

Loading

from datasets import load_dataset

ds = load_dataset("knguyennguyen/pattern2code", split="train")

ex = ds[0]
ex["standard_image"]        # PIL.Image — already decoded, no cast_column needed
ex["noise_image"]           # PIL.Image
print(ex["attribute_name"], ex["target_value"], "vs baseline", ex["bias_value"])
print(ex["code_snippet_context"])

Evaluation protocol

Prompt the model with standard_image (or noise_image) plus code_snippet_context:

You are doing a visual code fill-in-the-blank task. Given the webpage screenshot
and HTML snippet below, fill the blank token __ with ONLY the missing CSS/HTML
value. The goal is to predict the masked value to recreate the {pattern} in the
webpage design. Return the final answer in JSON format: {"answer":"<value>"}.
The blank must be a percentage value divisible by 10.

Score each prediction into three mutually exclusive outcomes:

• Correct — equals target_value
• Bias-aligned — equals bias_value (100%)
• Other error — neither

Citation

@inproceedings{nguyen2026pattern,
  title     = {Pattern Over Pixels: Measuring Pattern Completion Bias in
               Multimodal Code Generation},
  author    = {Nguyen, Khai-Nguyen and Chaparro, Oscar and Mastropaolo, Antonio},
  booktitle = {Proceedings of the 41st IEEE/ACM International Conference on
               Automated Software Engineering (ASE)},
  year      = {2026}
}

License & provenance

The seed webpages come from the Design2Code dataset, which is released under the Open Data Commons Attribution License (ODC-By) as it is built on the C4 corpus (also ODC-By). All perturbed HTML and screenshots here are derivative works and are redistributed under ODC-By with attribution. Copyright in the original website content remains with the respective site owners; the material is shared for research and reproducibility purposes.