ASO-Strategist

A fine-tuned Llama 3.2 3B model optimized for App Store Optimization (ASO) tasks, Given an app description, it generates:

• 🎯 Target Keywords - High-value search terms (5-8 keywords)
• 📝 Optimized Subtitle - Compelling 30-char App Store subtitle
• 📊 Search Visibility Score - 1-100 rating based on keyword potential

Model Details

Attribute	Value
Base Model	Llama-3.2-3B-Instruct-4bit
Quantization	4-bit
Fine-tuning	LoRA (16 layers, rank 8, alpha 16)
Framework	MLX
Training Data	1,000 synthetic ASO samples
Final Val Loss	0.320

Output Format

{
    "target_keywords": [
        "meditation app",
        "sleep stories",
        "stress relief",
        "mindfulness"
    ],
    "optimized_subtitle": "Calm your mind, sleep better",
    "search_visibility_score": 85,
    "reasoning": "Keywords target high-volume wellness searches..."
}

Quick Start

Installation

pip install mlx-lm

Inference

from mlx_lm import load, generate
from mlx_lm.sample_utils import make_sampler
import json

# Load model
model, tokenizer = load("fahidnasir/ASO-Strategist")
sampler = make_sampler(temp=0.3, top_p=0.9)

# System prompt
SYSTEM = """You are an expert App Store Optimization (ASO) specialist. Given an app description, provide:
1. Target Keywords: 5-8 high-value keywords/phrases for ASO
2. Optimized Subtitle: A compelling 30-character max subtitle
3. Search Visibility Score: 1-100 based on keyword potential

Respond in JSON format."""

# Build prompt (Llama 3.2 format)
prompt = f"""<|begin_of_text|><|start_header_id|>system<|end_header_id|>

{SYSTEM}<|eot_id|><|start_header_id|>user<|end_header_id|>

Analyze this app: A fitness tracking app with AI-powered workout recommendations and progress analytics<|eot_id|><|start_header_id|>assistant<|end_header_id|>

"""

# Generate
response = generate(model, tokenizer, prompt=prompt, max_tokens=512, sampler=sampler)

# Parse JSON
result = json.loads(response[response.find('{'):response.rfind('}')+1])
print(json.dumps(result, indent=2))

Example Output

Input:

"A meditation app with guided sleep stories, breathing exercises, and daily mindfulness reminders for stress relief"

Output:

{
    "target_keywords": [
        "meditation app",
        "guided sleep stories",
        "stress relief",
        "mindfulness reminders",
        "breathing exercises",
        "relaxation",
        "sleep stories",
        "mindfulness meditation"
    ],
    "optimized_subtitle": "Sleep better, stress less",
    "search_visibility_score": 85,
    "reasoning": "High-volume keywords in wellness/meditation category with good long-tail coverage"
}

Training Details

• Iterations: 500
• Batch Size: 4
• Learning Rate: 1e-4
• Max Sequence Length: 2048
• Hardware: Apple Silicon 64GB unified memory

Loss Curve

Iteration	Train Loss	Val Loss
50	0.396	0.376
100	0.341	0.355
200	0.317	0.334
300	0.290	0.325
500	0.260	0.320

Hardware Requirements

• Minimum: Apple Silicon Mac with 8GB unified memory
• Recommended: M1 Pro/Max/Ultra or M2/M3/M4 with 16GB+

Limitations

• Trained on synthetic data; real-world ASO may vary
• English language only
• Subtitle recommendations may occasionally exceed 30 characters
• Keywords reflect general patterns, not real-time App Store data

Use Cases

• 📱 App Developers: Quick keyword research for new app listings
• 📈 Marketing Teams: Baseline ASO analysis and subtitle ideation
• 🔍 ASO Professionals: Automated first-pass keyword suggestions
• 📚 Learning: Understanding ASO principles and keyword selection

License

This model inherits the Llama 3.2 Community License.

Built with Llama.

Citation

@misc{aso-strategist-2025,
  author = {Fahid Nasir},
  title = {ASO-Strategist: Fine-tuned Model for App Store Optimization},
  year = {2025},
  publisher = {Hugging Face},
  url = {https://huggingface.co/fahidnasir/ASO-Strategist}
}

Acknowledgments

• Meta AI for Llama 3.2
• MLX Team for the Apple Silicon ML framework
• Hugging Face for model hosting