summykai/tailwind-v4-sft-mix-001

Tailwind-SFT: Instruction-Tuning Dataset for Tailwind CSS v4

Dataset: summykai/tailwind-v4-sft-mix-001
License: MIT (Tailwind CSS)

Dataset Description

Tailwind-SFT is a 5,610-example instruction-tuning dataset mapping natural language UI intents to validated Tailwind CSS v4 utility classes and React/JSX snippets. Each example is compiler-verified: all 1,122 utility classes were validated against @tailwindcss/cli (v4.1.18) to eliminate hallucination and ensure syntactic correctness.

The dataset addresses a critical gap in existing CSS datasets—lack of ground-truth validation—by establishing a compiler-enforced baseline for grounded utility-class generation. This makes it suitable for training domain-specific models that interface with modern CSS frameworks.

Task Types

• choose_classes (3,744 examples): Return optimized class strings for UI components
• react_snippet (1,496 examples): Generate production-ready JSX with Tailwind classes
• explain_choice (370 examples): Explain class selection rationale given anchor utilities

Technical Specifications

Hardware: NVIDIA RTX 4070 Ti (12GB)
Peak VRAM: 3.91 GB
Generation Rate: 3.08 examples/sec

Model: SmolLM3-3B 4-bit NF4
Optimizations: Flash Attention 2, torch.compile(mode="max-autotune"), prefetch batching

Validation: 100% pass rate against Tailwind v4 compiler
Split: Anchor-class stratified (5,045 train / 565 validation)
Metadata: 11 structured fields per example for curriculum learning

Dataset Structure

{
  "messages": [
    {"role": "user", "content": "Suggest Tailwind v4 classes for a primary button"},
    {"role": "assistant", "content": "Use these utilities: `flex px-4 py-2 bg-blue-600`..."}
  ],
  "metadata": {
    "anchor_class": "flex",
    "category": "layout",
    "archetype": "button",
    "tailwind_version": "4.1.18",
    "classes_used": ["flex", "px-4", "py-2", "bg-blue-600"],
    "template_type": "choose_classes",
    "has_code": false,
    "license": "MIT (Tailwind CSS)"
  }
}

Usage

from datasets import load_dataset

ds = load_dataset("summykai/tailwind-v4-sft-mix-001")

# Filter by component archetype
button_ds = ds.filter(lambda ex: ex["metadata"]["archetype"] == "button")

# Curriculum learning by category
layout_ds = ds.filter(lambda ex: ex["metadata"]["category"] == "layout")

Benchmarks & Expected Performance

Models trained on this dataset should demonstrate:

• Class Accuracy >95% of generated classes are valid Tailwind v4 utilities
• Correct variant syntax (hover:, focus:, sm:, md:)
• Bundle coherence: Logically grouped utilities (spacing + typography + color)
• Zero hallucination: No fabricated classes like bg-blue-450

Generation Pipeline

1. Catalog Build: 1,122 candidate classes from structured generation rules
2. CLI Validation: All classes verified via Tailwind v4 compiler
3. Template Seeding: High-signal archetype-driven examples (7 component types)
4. Paraphrasing: SmolLM3 with 85% user / 45% assistant rewrite probability
5. Quality Control: Duplicate removal, class validity filtering

Citation

@dataset{tailwind_sft_2025,
  title={Tailwind-SFT: High-Signal Instruction-Tuning Dataset for Tailwind CSS v4},
  author={SummyKai},
  year={2025},
  url={https://huggingface.co/datasets/summykai/tailwind-v4-sft-mix-001}
}

Limitations

• Coverage limited to 1,122 high-frequency utilities (not exhaustive)
• Component archetypes restricted to 7 common patterns
• Paraphrasing may occasionally over-generalize technical details

Contact

For issues or questions: GitHub Issues