NOTAM Adapter Training

Train a Foundation Models LoRA adapter to extract structured runway performance data from raw NOTAM text.

Quick Start

# 1. Setup environment
./setup.sh

# 2. Prepare training data
./prepare_data.sh

# 3. Train the adapter
./train_adapter.sh

# 4. Upload to App Store Connect
./upload_adapter.sh

Overview

This pipeline trains a custom LoRA adapter for Apple's Foundation Models framework to parse NOTAMs and extract runway performance data.

[NOTAM API] → [Filter] → [Label with Claude] → [Train] → [Export] → [Upload]

Prerequisites

• macOS with Apple Silicon (32GB+ RAM recommended)
• Python 3.11 (via pyenv)
• Apple's adapter training toolkit
• Anthropic API key (for labeling)
• Apple Developer Program membership (for hosting)

Setup

1. Install Apple's Toolkit

Download from Apple Developer:

unzip adapter_training_toolkit_v26_0_0.zip

2. Setup Python Environment

./setup.sh

3. Configure Environment

cp .env.example .env
# Edit .env with your credentials

Required variables:

• ANTHROPIC_API_KEY - For Claude-based labeling
• TOOLKIT_PATH - Path to Apple's toolkit
• ASC_ISSUER_ID, ASC_KEY_ID, ASC_PRIVATE_KEY_PATH - App Store Connect API credentials
• APP_APPLE_ID - Your app's Apple ID (numeric)

Data Preparation

./prepare_data.sh

Or run individual steps:

1. python download_all_notams.py - Download NOTAMs from API
2. python filter_relevant_notams.py - Filter runway-relevant NOTAMs
3. python generate_silver_labels.py - Label with Claude
4. python review_tool.py --low-confidence - Review low-confidence labels
5. python fix_silver_labels.py - Apply automatic fixes
6. python format_training_data.py - Format for training

Training

./train_adapter.sh

Options:

• --export-only - Export existing checkpoint without training

Environment variables:

Variable	Default	Description
TOOLKIT_PATH	required	Apple toolkit location
EPOCHS	5	Training epochs
BATCH_SIZE	4	Batch size (reduce if OOM)
LEARNING_RATE	1e-3	Learning rate

For memory issues:

BATCH_SIZE=1 ./train_adapter.sh

Upload

./upload_adapter.sh

Options:

• --dry-run - Validate without uploading

The script uses the App Store Connect API to:

1. Create/find an asset pack
2. Create a new version
3. Upload the adapter
4. Commit and verify processing

Required Entitlement

Request the Foundation Models Framework Adapter Entitlement from Apple before shipping:

1. Go to Apple Developer Account
2. Certificates, Identifiers & Profiles → Identifiers
3. Select your App ID → Request entitlement

Directory Structure

├── setup.sh               # Environment setup
├── prepare_data.sh        # Data preparation pipeline
├── train_adapter.sh       # Training script
├── upload_adapter.sh      # Upload to App Store Connect
├── requirements.txt       # Python dependencies
├── .env.example           # Environment template
├── data/                  # Training data (gitignored)
├── checkpoints/           # Model checkpoints (gitignored)
└── exports/               # Exported adapters (gitignored)

Data Schema

Training Format

[
  {"role": "user", "content": "Extract runway data from this NOTAM...\n\nRWY 28L THR DSPLCD 500FT"},
  {"role": "assistant", "content": "{\"airportID\":\"KSFO\",\"runway\":\"28L\",\"takeoffShortening\":500,...}"}
]

Extraction Fields

Field	Type	Description
airportID	string	ICAO airport code
runway	string?	Runway designator
runwayClosed	bool	Runway fully closed
takeoffShortening	number?	Takeoff distance reduction
landingShortening	number?	Landing distance reduction
TORA, TODA, LDA	number?	Declared distances
obstacleHeight	number?	Obstacle height AGL
contaminations	array	Surface contaminations

References

• Foundation Models adapter training
• Background Assets framework
• App Store Connect API

License

MIT License - see LICENSE