README.md

---
license: cc-by-4.0
task_categories:
- text-generation
---

# 🧠 CC-HARD: A Challenging Dataset for Design-to-Code Generation

[📄 Paper on arXiv](https://arxiv.org/pdf/2508.03560)
[📄 Paper on ACM](https://arxiv.org/pdf/2508.03560)

**CC-HARD** is a challenging benchmark dataset introduced in the KDD 2025 paper *LaTCoder: Converting Webpage Design to Code with Layout-as-Thought*.
It was specifically designed to evaluate layout fidelity in **webpage design-to-code generation**.

The dataset consists of **128 webpage screenshots** and their corresponding **HTML/CSS code**, manually curated from the Common Crawl corpus. Unlike prior datasets, CC-HARD emphasizes:

- **Deep DOM hierarchies**
- **Visually complex and diverse layouts**
- **High tag density and structural variability**

LaTCoder uses CC-HARD to demonstrate that existing MLLM-based design-to-code approaches struggle with layout preservation when faced with real-world webpages. This dataset provides a strong testbed for evaluating both general-purpose MLLMs and layout-aware code generation systems.

---

## 📦 Dataset Details

Each example in **CC-HARD** consists of:

- `image`: A high-resolution PNG screenshot of a real-world webpage design  
- `text`: The corresponding HTML/CSS code used to render that design


## Dataset Comparison: CC-HARD vs. Design2Code-HARD

We compare **CC-HARD** with **Design2Code-HARD** dataset across multiple structural and content dimensions.
The following table summarizes the key statistics, as reported in the *LaTCoder* paper (Table 2):

| Metric                   | Design2Code-HARD | CC-HARD |
|--------------------------|------------------|---------|
| Number of Samples        | 80               | 128     |
| Avg. Total Length (tokens) | 8,900 ± 2,399   | 8,416 ± 2,190 |
| Avg. Text Length (tokens)  | 3,554 ± 2,820   | 969 ± 762     |
| Avg. HTML Tags           | 251 ± 232        | 274 ± 66      |
| Avg. DOM Depth           | 10 ± 4           | 16 ± 3        |
| Avg. Unique Tags         | 23 ± 5           | 27 ± 5        |


### Key Differences and Analysis

As discussed in the paper:

- While **overall token lengths** of samples are similar across datasets, **text content** in Design2Code-HARD is much longer, meaning it emphasizes textual richness.
- In contrast, **CC-HARD’s token budget is dominated by HTML structure**, not content, making it harder for models to learn from context or infer layout from semantic cues.
- **CC-HARD samples are structurally deeper** (DOM depth 16 vs. 10), with **more HTML tags** and **greater tag diversity**, increasing layout reasoning complexity.
- CC-HARD features a **larger number of nested layout blocks**, which amplifies the difficulty for MLLMs in preserving spatial relationships and hierarchy during code generation.

As a result, models that perform well on Design2Code-HARD often struggle on CC-HARD — a trend clearly shown in the benchmark results.
This highlights the **increased layout sensitivity and real-world difficulty** embedded in CC-HARD, making it a more suitable testbed for evaluating layout-aware design-to-code systems.

---

## 🧾 Citation

```bibtex
@inproceedings{gui2025latcoder,  
    author = {Gui, Yi and Li, Zhen and Zhang, Zhongyi and Wang, Guohao and Lv, Tianpeng and Jiang, Gaoyang and Liu, Yi and Chen, Dongping and Wan, Yao and Zhang, Hongyu and Jiang, Wenbin and Shi, Xuanhua and Jin, Hai},
    title = {LaTCoder: Converting Webpage Design to Code with Layout-as-Thought},
    year = {2025},
    isbn = {9798400714542},
    publisher = {Association for Computing Machinery},
    address = {New York, NY, USA},
    url = {https://doi.org/10.1145/3711896.3737016},
    doi = {10.1145/3711896.3737016},
    booktitle = {Proceedings of the 31st ACM SIGKDD Conference on Knowledge Discovery and Data Mining V.2},
    pages = {721–732},
    numpages = {12},
    keywords = {code generation, design to code, ui automation},
    location = {Toronto ON, Canada},
    series = {KDD '25}  
}