syaher/web-builder-dataset

Web Builder Dataset — Fine-tuning untuk LLM

Dataset untuk fine-tune model pada pembangunan web builder (WeWeb clone) menggunakan Nuxt 4 + Tailwind CSS 4 + Cloudflare.

---

📊 Dataset Overview

Metric	Value
Total examples	500
Train	400 (80%)
Eval	100 (20%)
Format	ChatML JSONL
Language	UI: Bahasa Malaysia, Code: English

---

🎯 Fokus Dataset

Kategori	Bilangan Contoh
Nuxt 4 Pages & Config	~50
API Handlers (Drizzle)	~80
Builder Components (Blocks)	~120
Auth & Middleware	~40
Dashboard & Admin	~60
Payment & Billing	~40
Landing Pages	~60
Utilities & Composables	~50

---

🏗️ Stack yang Dijarab

• Framework: Nuxt 4.5.2
• CSS: Tailwind CSS 4
• Database: Drizzle ORM (SQLite/Turso)
• Cloud: Cloudflare (Pages, Workers, R2)
• Payment: ToyyibPay
• Drag & Drop: @dnd-kit
• Icons: @nuxt/icon (Iconify)

---

📁 Struktur Fail

dataset/
├── train.jsonl     # 400 contoh latihan
├── eval.jsonl      # 100 contoh evaluation
└── README.md       # dokumen ini

---

🔧 Cara Penggunaan

1. LLaMA-Factory

# Clone framework
git clone https://github.com/hiyouga/LLaMA-Factory.git
cd LLaMA-Factory

# Upload dataset
scp dataset/train.jsonl root@<instance>:/root/LLaMA-Factory/data/web_builder.jsonl

# Edit data_config.yaml
# - dataset_name: web_builder
# - file_name: web_builder.jsonl

# Train (QLoRA 4-bit)
llamafactory-cli train \
  --model_name_or_path Qwen/Qwen2.5-7B-Instruct \
  --dataset web_builder \
  --finetuning_type lora \
  --quantization_bit 4 \
  --template chatml \
  --output_dir ./output/web_builder

2. Unsloth (Faster training)

# Install unsloth
pip install unsloth

# Run training
python train.py \
  --model Qwen/Qwen2.5-7B-Instruct \
  --dataset data/web_builder.jsonl \
  --max_seq_length 2048 \
  --per_device_train_batch_size 4 \
  --learning_rate 2e-4 \
  --num_train_epochs 3 \
  --lora_r 16 \
  --lora_alpha 32

3. Axolotl

# config.yaml
model: Qwen/Qwen2.5-7B-Instruct
dataset: web_builder
dataset_format: chatml
sequence_len: 2048

lora:
  r: 16
  alpha: 32
  dropout: 0.05
  target_modules: all

train:
  num_epochs: 3
  batch_size: 4
  learning_rate: 2e-4

---

📝 Format Data

Setiap baris adalah JSON object dengan messages array:

{
  "messages": [
    {"role": "system", "content": "Kau developer Nuxt 4 + Tailwind CSS 4 + Cloudflare..."},
    {"role": "user", "content": "Buat API handler untuk list pages"},
    {"role": "assistant", "content": "export default defineEventHandler..."}
  ]
}

System Prompt (Fixed)

Kau developer Nuxt 4 + Tailwind CSS 4 + Cloudflare. UI strings Bahasa Malaysia, kod English. Monorepo pnpm workspace.

---

⚙️ Training Hyperparameters (Cadangan)

Parameter	Value
Model base	Qwen2.5-7B-Instruct
Max sequence	2048
Batch size	4
Learning rate	2e-4
Epochs	3
LoRA r	16
LoRA alpha	32
Quantization	4-bit

---

🔍 Evaluation

Run inference pada eval.jsonl untuk test model:

llamafactory-cli inference \
  --model_name_or_path ./output/web_builder \
  --template chatml \
  --messages "Buat API handler untuk create page"

---

📦 Download Dataset

# Clone
git clone https://github.com/your-repo/web-builder-dataset.git
cd web-builder-dataset

# Verify
wc -l dataset/train.jsonl  # Should be 400
wc -l dataset/eval.jsonl   # Should be 100

---

🏷️ Tags & Categories

• nuxt4
• tailwind4
• cloudflare
• drizzle
• web-builder
• landing-page
• drag-drop
• saas
• malay-ui

---

📜 License

MIT — Free untuk penggunaan komersial dan bukan-komersial.

---

Dataset dijana secara automatik berdasarkan pattern pembangunan web builder. Semua kod adalah contoh training, verify sebelum penggunaan production.