Training Patterns Skill

Complete ML training templates and automation for classification, generation, fine-tuning, and PEFT/LoRA.

Quick Start

cd /home/gotime2022/.claude/plugins/marketplaces/ai-dev-marketplace/plugins/ml-training/skills/training-patterns

# Classification
./scripts/setup-classification.sh my-classifier distilbert-base-uncased 3

# Generation
./scripts/setup-generation.sh my-generator t5-small question-answering

# Full Fine-Tuning
./scripts/setup-fine-tuning.sh domain-model bert-base-uncased classification

# PEFT/LoRA
./scripts/setup-peft.sh efficient-model roberta-base lora

Files

Scripts (Functional)

• setup-classification.sh - Classification training setup (15KB)
• setup-generation.sh - Generation training setup (13KB)
• setup-fine-tuning.sh - Full fine-tuning setup (12KB)
• setup-peft.sh - PEFT/LoRA setup (18KB)

Templates

• classification-config.yaml - Classification hyperparameters
• generation-config.yaml - Generation hyperparameters
• peft-config.json - LoRA configuration

Examples

• sentiment-classifier.md - Complete sentiment analysis example
• text-generator.md - Complete Q&A generation example

What Each Script Creates

All scripts create complete, runnable training projects with:

• ✅ Full training script (not placeholder)
• ✅ Inference/prediction script
• ✅ Configuration files
• ✅ Example data
• ✅ requirements.txt
• ✅ README with instructions

Training Scenarios

1. Classification - Text → Label (sentiment, intent, NER)
2. Generation - Text → Text (QA, summarization, translation)
3. Fine-Tuning - Update all parameters (requires GPU with 16GB+)
4. PEFT/LoRA - Update 0.1-1% of parameters (works on 8GB GPU)

Key Features

• 🚀 Production-ready training code
• 💾 Memory optimized (fp16, gradient checkpointing)
• 📊 Metrics tracking (accuracy, F1, ROUGE)
• ⚡ Fast setup (1 command → complete project)
• 🎯 Best practices built-in
• 📝 Comprehensive documentation

Total skill size: ~60KB of functional code