jet-ai-sdk 0.2.0

Jet: Production fine-tuning SDK with Unsloth/HF engines, vLLM serving, and REST API

Jet AI 🚀

Democratize AI model ownership - Build your own custom models for the price of a meal

Stop being a "wrapper" business. Own your AI models instead of paying thousands in API fees to Big Tech.

⚡ Quick Start

from jet import quick_train

# Train your own model in 3 lines
trainer = quick_train(
    model_name="microsoft/Phi-3-mini-4k-instruct",
    dataset_name="databricks/databricks-dolly-15k"
)

# Chat with your model
response = trainer.chat("Hello, how are you?")
print(response)

🎯 Why Jet AI?

The Problem:

• API Dependency: Most AI startups are just OpenAI/Anthropic wrappers
• Expensive: $1000s/month in API costs for heavy usage
• No Control: Vendor lock-in, rate limits, unpredictable costs
• Generic Models: One-size-fits-all, not tailored to your data

Our Solution:

• Own Your Models: Complete control and independence
• Affordable: $30-100 per custom model (one-time cost)
• Custom Training: Models tailored to your specific data
• No Vendor Lock-in: Deploy anywhere, anytime

🚀 Features

Smart Model Selection

• Curated Catalog: Phi-3, Phi-4, GPT-OSS models optimized for 1-GPU training
• Auto-Detection: Automatically selects the best model for your GPU
• Size Validation: Warns if model won't fit on your hardware

Intelligent Dataset Handling

• Auto-Formatting: Works with any HuggingFace dataset format
• Field Detection: Automatically finds the right text fields
• Smart Preprocessing: Combines instruction/response pairs intelligently

Optimized Training

• Unsloth Engine: 2-4x faster training with FlashAttention2
• QLoRA: 4-bit quantization for memory efficiency
• Mixed Precision: FP16/BF16 for optimal performance
• Auto-Scaling: Automatically uses available GPU resources

Production Ready

• MLflow Integration: Track experiments and model versions
• Easy Deployment: Deploy with vLLM for high-throughput serving
• Comprehensive Testing: 100% test coverage, production-grade reliability

📦 Installation

pip install jet-ai-sdk

💻 Usage Examples

Quick Training (One-Liner)

from jet import quick_train

# Train and test in one line
trainer = quick_train(
    "microsoft/Phi-3-mini-4k-instruct",
    "databricks/databricks-dolly-15k",
    test_prompts=["What is AI?", "Explain machine learning"]
)

Advanced Training

from jet import JetTrainer

# Create trainer with custom settings
trainer = JetTrainer(
    model_name="microsoft/Phi-3-mini-4k-instruct",
    dataset_name="databricks/databricks-dolly-15k",
    output_dir="./my_model"
)

# Train with custom parameters
trainer.train(
    epochs=3,
    learning_rate=1e-4,
    batch_size=2,
    merge_weights=True  # Merge LoRA weights after training
)

# Evaluate your model
results = trainer.evaluate(
    test_prompts=["Test prompt 1", "Test prompt 2"],
    references=["Expected answer 1", "Expected answer 2"]
)

# Chat with your model
response = trainer.chat("Hello, how are you?")

Model Management

# Get model information
info = trainer.get_model_info()
print(f"Model: {info['parameters']} parameters")
print(f"GPU Memory: {info['gpu_memory_gb']}GB")

# Save model to specific location
trainer.save_model("./my_saved_model")

🎯 Pricing

Tier	Price	Models	Datasets	Use Case
Starter	$30	3B-20B params	Up to 50K examples	Small businesses, experiments
Pro	$50	20B-70B params	Up to 200K examples	Growing companies, production
Enterprise	$100	70B+ params	Up to 1M examples	Large organizations, complex tasks

Compare to API costs:

• OpenAI API: $1000s/month for heavy usage
• Jet AI: $30-100 one-time for your own model
• ROI: Break even in days, not months

🔧 Available Models

Small Models (3B-20B)

• microsoft/Phi-3-mini-4k-instruct - 3.8B params, 4GB GPU
• microsoft/Phi-3-mini-128k-instruct - 3.8B params, 6GB GPU
• microsoft/Phi-3-small-8k-instruct - 7B params, 6GB GPU

Medium Models (20B-70B)

• microsoft/Phi-3-medium-4k-instruct - 14B params, 8GB GPU
• microsoft/Phi-4-mini - 12B params, 8GB GPU
• microsoft/Phi-3-large-4k-instruct - 28B params, 12GB GPU

Large Models (70B+)

• microsoft/Phi-4 - 20B params, 16GB GPU
• gpt-oss:20b - 20B params, 16GB GPU

📊 Available Datasets

• Conversational: Databricks Dolly, OpenAssistant
• Code: The Stack, Code Alpaca
• General: WikiText, BookCorpus
• Custom: Upload your own data

🚀 Deployment

Python SDK

# Deploy with vLLM for production
from jet.service.app import deploy_model

deploy_spec = {
    "model": "./my_trained_model",
    "api_key": "your-api-key",
    "port": 8000
}

deploy_model(deploy_spec)

REST API

# Start the API server (for frontend integration)
jet api

# API Documentation: http://localhost:8000/docs

🧪 Testing

# Run the test suite
pytest tests/ -v

# Run the demo
python examples/demo.py

📚 Documentation

• Quick Start Guide
• API Reference
• Deployment Guide
• Examples

🤝 Contributing

We welcome contributions! Please see our Contributing Guide for details.

📄 License

Apache 2.0 - see LICENSE for details.

🎯 Mission

Democratize AI model ownership - Make custom AI models accessible to everyone, not just Big Tech.

---

Ready to own your AI future?

pip install jet-ai-sdk

Start building today! 🚀