Datasets:
metadata
pretty_name: codey-bench
license: cc-by-nc-4.0
language:
- ko
- en
task_categories:
- text-generation
tags:
- structured-output
- json
- web-design
size_categories:
- n<1K
configs:
- config_name: ko
default: true
data_files:
- split: test
path: data/ko/test.jsonl
- config_name: en
data_files:
- split: test
path: data/en/test.jsonl
- config_name: ko-synth
data_files:
- split: test
path: data/ko-synth/test.jsonl
codey-bench
Each row is a product brief for a web page (Markdown). The model receives the brief and must produce one Codey Design Document (CDD) as JSON, the page's structured design: sections, blocks and theme. A reference document is included for each brief.
system_prompt.txt is the system message for every row. It states the task and the CDD
format, and ends with "Answer with one JSON object and nothing else."
Subsets
| subset | rows | language | briefs and references |
|---|---|---|---|
ko (default) |
20 | Korean | written by product planners |
en |
99 | English | synthesized with gpt-4.1 |
ko-synth |
98 | Korean | Korean versions of en, same ids |
Fields
| field | description |
|---|---|
id |
document id; en and ko-synth share ids |
lang |
ko or en |
domain |
business domain, e.g. hair-salon, agency |
brief |
the Markdown brief, used as the user message |
reference |
the reference CDD as a JSON string |
gold_schema_valid |
the reference conforms to the CDD schema (always true) |
gold_gate_errors |
number of rendering-rule violations in the reference; 0 for every ko row. Synthesized references keep asset metadata keys and section links that the renderer rejects |
Usage
from datasets import load_dataset
from huggingface_hub import hf_hub_download
ds = load_dataset("AISeedCorp/codey-bench", "ko", split="test")
system = open(hf_hub_download("AISeedCorp/codey-bench", "system_prompt.txt", repo_type="dataset")).read()
messages = [{"role": "system", "content": system},
{"role": "user", "content": ds[0]["brief"]}]
License
CC BY-NC 4.0.