Auroraventures/cipher-awwwards-sft25
收藏Hugging Face2026-04-19 更新2026-04-26 收录
下载链接：
https://hf-mirror.com/datasets/Auroraventures/cipher-awwwards-sft25
下载链接
链接失效反馈官方服务：
资源简介：
---
license: cc-by-nc-4.0
language:
- en
pretty_name: Cipher Awwwards SFT 2.5 + Real v1
size_categories:
- 1K<n<10K
task_categories:
- text-generation
tags:
- code
- html
- css
- javascript
- awwwards
- creative-web
- cipher
- kin
- three.js
- gsap
- lenis
- scrolltrigger
- splittext
- sft
- chat-format
- gemma-4
- web-design
configs:
- config_name: synthetic-awwwards
data_files:
- split: train
path: awwwards-stage25-sft.jsonl
- config_name: synthetic-awwwards-v2
data_files:
- split: train
path: awwwards-stage25-v2-sft.jsonl
- config_name: real-scraped-v1
data_files:
- split: train
path: cipher-real-v1-sft.jsonl
- config_name: gold-curated
data_files:
- split: train
path: awwwards-gold.jsonl
- config_name: grpo-diverse
data_files:
- split: train
path: awwwards-grpo-diverse.jsonl
---
# Cipher — Awwwards SFT 2.5 + Real v1 🦑
> *The training fuel for Kin's creative-web generator, AND the retrieval corpus for [Kraken RAG](https://github.com/kr8tiv-ai/kr8tiv-runtime-truth-contracts/tree/main/kraken_rag). 96 real Awwwards Site-of-the-Day winners + ~1,200 records from official motion-library repositories.*
## Two ways this dataset is used
1. **As a retrieval corpus for [Kraken RAG](https://github.com/kr8tiv-ai/kr8tiv-runtime-truth-contracts/tree/main/kraken_rag) ⭐ (the production path).** The `awwwards-gold.jsonl` file contains 96 structured records of real Awwwards SOTD winners — tags, tech stack, motion libs, CSS features, section counts, HTML size — which `kraken_rag` indexes with TF-IDF. A creative brief retrieves the 3 most stylistically similar winners and they're packed as in-context examples for a frontier model. This is the architecture Vercel's v0 is built around.
2. **As a supervised fine-tuning corpus for the `cipher-sft25-*` checkpoint family (an experiment).** The `real-scraped-v1` split (741 records from official Three.js / Motion One / GSAP examples) was used to fine-tune `Auroraventures/cipher-sft25-real-merged`. This experiment confirmed what the repo's own [CRITICAL-ASSESSMENT.md](https://github.com/kr8tiv-ai/kr8tiv-training/blob/main/data/awwwards/distilled/CRITICAL-ASSESSMENT.md) had warned on 2026-04-15: a 31 B local fine-tune on library examples converges on syntactically correct but aesthetically generic output. The weights remain useful as an offline scaffolder; for creative output, use Kraken RAG.
It ships **five splits** so every use case is reproducible.
---
## Splits at a glance
| Config | File | Records | Size | Purpose |
|---|---|---:|---:|---|
| `real-scraped-v1` ⭐ | `cipher-real-v1-sft.jsonl` | 741 | 5.66 MB | v3 SFT — real scraped source code (recommended) |
| `synthetic-awwwards-v2` | `awwwards-stage25-v2-sft.jsonl` | 96 | 739 KB | v2 synthetic SFT |
| `synthetic-awwwards` | `awwwards-stage25-sft.jsonl` | 288 | 1.88 MB | v1 synthetic SFT |
| `gold-curated` | `awwwards-gold.jsonl` | ~50 | 147 KB | hand-ranked "gold" exemplars |
| `grpo-diverse` | `awwwards-grpo-diverse.jsonl` | ~100 | 157 KB | GRPO-candidate prompt pool |
Each record is a Gemma-4 chat-format triple:
```json
{
"messages": [
{"role": "system", "content": "You are Cipher, the Code Kraken. Emit complete single-file HTML — no markdown fences, no preamble."},
{"role": "user", "content": "Build an Awwwards-quality hero with WebGL fluid and GSAP split-text entry."},
{"role": "assistant", "content": "<!DOCTYPE html>…"}
]
}
```
---
## ⭐ The `real-scraped-v1` split
The breakthrough split. Four canonical sources, normalized into the Cipher chat format:
| Source | Records | What it teaches |
|---|---:|---|
| [`mrdoob/three.js/examples`](https://github.com/mrdoob/three.js/tree/master/examples) | 578 | Canonical Three.js patterns — shaders, postprocessing, particles, raycasting, physics |
| [`motiondivision/motion/dev`](https://github.com/motiondivision/motion) | 148 | Motion One idioms transplanted to vanilla DOM |
| [freefrontend.com GSAP gallery](https://freefrontend.com/gsap-examples/) | 63 | ScrollTrigger timelines, SplitText chains, SVG morphs |
| [aura.build](https://aura.build/) shells | ≤ 998 | Modern CSS scaffolding, typography tokens, dark-mode palettes |
Every record is real, hand-checked code. No LLM-generated synthetic content. No Tailwind CDN. No `lenis.stop()` misuse.
---
## Load it
```python
from datasets import load_dataset
ds = load_dataset(
"Auroraventures/cipher-awwwards-sft25",
name="real-scraped-v1",
split="train",
)
print(ds[0]["messages"][-1]["content"][:500])
```
Or with Pandas / Polars:
```python
import polars as pl
df = pl.read_ndjson("hf://datasets/Auroraventures/cipher-awwwards-sft25/cipher-real-v1-sft.jsonl")
```
---
## Why real > synthetic (for creative code)
Synthetic SFT (distilling one instruct-tuned LM with another) converges on the teacher's mean aesthetic. For Awwwards-grade web design, the mean is *boring* — and the long tail is where the award-winning sites live.
v2 (`synthetic-awwwards`) suffered **template collapse**: identical hero layouts across divergent prompts, repeated palettes, repeated section ordering. The v3 fix was `real-scraped-v1` — authentic code from the sources that actually shipped.
---
## Intended use
- ✅ SFT / SimPO / GRPO / KTO training of creative-code LMs
- ✅ Evaluation of web-design code generation
- ✅ Research on diversity, slop, and mode-collapse in code models
- ❌ Commercial redistribution of the raw code (see license)
---
## Licensing & attribution
- **Dataset license:** CC-BY-NC-4.0
- **Source code:** each upstream source retains its own license (mostly MIT or permissive). The `real-scraped-v1` split is a transformed, aggregated research corpus and should not be used to bypass upstream licenses for commercial redistribution. When in doubt, link back to the original source.
- **Cipher checkpoints trained on this dataset:** CC-BY-NC-4.0.
---
## Citation
```bibtex
@misc{cipher-awwwards-sft25-2026,
title = {Cipher Awwwards SFT 2.5 + Real v1: a creative-web supervision corpus},
author = {Matt Haynes and Aurora Ventures},
year = {2026},
howpublished = {\url{https://huggingface.co/datasets/Auroraventures/cipher-awwwards-sft25}},
}
```
---
## Changelog
| Date | Split | Notes |
|---|---|---|
| 2026-04-14 | `gold-curated`, `grpo-diverse` | Initial hand-ranked gold set + GRPO prompt pool |
| 2026-04-15 | `synthetic-awwwards` | v1 synthetic SFT (288 records) |
| 2026-04-16 | `synthetic-awwwards-v2` | v2 synthetic SFT (96 records) — retired after template collapse |
| 2026-04-18 | `real-scraped-v1` ⭐ | Real-code SFT breakthrough (741 records, 4 canonical sources) |
---
*Curated with 🦑 by [Aurora Ventures](https://huggingface.co/Auroraventures). Trains the [Cipher](https://huggingface.co/Auroraventures/cipher-sft25-real-merged) generator inside the [Kin](https://github.com/kr8tiv-ai/Kin) runtime.*
提供机构：
Auroraventures
