---
pretty_name: AUV-BENCH Data
task_categories:
- other
tags:
- ui
- web-ui
- aesthetics
- benchmark
- multimodal
- vision-language
size_categories:
- 1K<n<10K
license: other
---

# AUV-BENCH Data

**AUV-BENCH** is a public benchmark dataset for evaluating multimodal models on
four UI-aesthetics capabilities over executable webpages: **scoring**,
**diagnosis**, **repair**, and **text-to-UI generation**.

## Tasks

| Directory | Task | Contents | Scale |
|---|---|---|---|
| `task1/` | Aesthetic Scoring | Webpage screenshots + human 8-dimension aesthetic MOS ratings (`human_gt_cleaned_mos.csv`) | 1,395 samples |
| `task2/` | Aesthetic Diagnosis | Paired clean/degraded screenshots + injected-defect ground truth | 695 pairs (660 formal) |
| `task3/` | Aesthetic Repair | Degraded webpage HTML + defect specs + 25 pinned render fonts | 660 samples |
| `task4/` | Text-to-UI Generation | Generation queries + Top-K visual assets (image + caption) | 1,395 queries |

## Directory structure

All data ships as a single archive, `auv_bench_data_v1.tar.gz`, which extracts
to the following layout:

```text
auv_bench_data_v1/
├── MANIFEST.txt                 # per-task file/sample counts
├── task1/  manifest.jsonl  images/  human_gt_cleaned_mos.csv
├── task2/  manifest.jsonl  images/
├── task3/  cohort.json  defect_specs.jsonl  pages/  render_fonts/fonts/
└── task4/  queries.jsonl  cohort.json  freeze.json  inputs/
```