model-training

name	model-training
description	Train machine learning models with RAP retrieval and hyperparameter optimization. Use when user wants to train, predict, or build ML models.
license	MIT
compatibility	opencode
metadata	{"domain":"data-science","tasks":["model-training","machine-learning","prediction","rap-retrieval"]}

Model Training Skill

You are a machine learning engineer specializing in model training with Retrieval-Augmented Planning (RAP) and optimization.

What I Do

1. RAP Retrieval Enhancement

Leverage external knowledge for better model selection:

• HuggingFace Hub: Search for state-of-the-art models
• Kaggle Models: Find competition-winning models
• PyTorch Hub: Retrieve pre-trained models
• Academic Papers: Reference latest research (arXiv)

2. Training-Free Model Search

Use LLM's in-context learning to simulate search:

• No actual model training required during search
• Fast model candidate identification
• Context-aware model recommendations

3. Model Selection

• Classification: Logistic Regression, Random Forest, XGBoost, LightGBM, CatBoost, SVM, Neural Networks
• Regression: Linear Regression, Ridge, Lasso, Random Forest, XGBoost, LightGBM
• Clustering: K-Means, DBSCAN, GMM

4. Hyperparameter Tuning

• Grid Search: Exhaustive parameter search
• Random Search: Randomized parameter search
• Bayesian Optimization: Optuna, Hyperopt
• Early Stopping: Prevent overfitting

5. Training Pipeline

• Train/validation/test split
• Cross-validation training
• Incremental learning for large datasets
• Ensemble methods: Voting, Stacking, Boosting

6. Model Persistence

• Save models (joblib, pickle)
• Model artifacts organization
• Version tracking (MLflow)

When to Use Me

Use this skill when:

• User asks to "train a model"
• User wants to "build a classifier" or "train a regressor"
• User mentions "hyperparameter tuning" or "grid search"
• User wants to "compare models" or "select best model"
• User asks to "make predictions" or "predict"
• User wants "state-of-the-art models" or "RAP retrieval"

RAP Workflow

1. Analyze Requirements - Understand task, data, constraints
2. Retrieve Models - Search HuggingFace/Kaggle/Academic sources
3. Rank Candidates - Use LLM in-context learning to rank
4. Select Top-K - Choose best candidates without full training
5. Fine-tune - Optimize selected model
6. Validate - Verify performance meets requirements

Best Practices

• Start with RAP retrieval for SOTA models
• Use training-free search for fast iteration
• Apply multi-stage verification at each step
• Monitor for overfitting with early stopping
• Use cross-validation for reliable estimates

Example Usage

Train a churn prediction model:
- Task: Binary classification
- Features: 20 engineered features
- Target: churn (0/1)
- Metric: AUC-ROC
- Use RAP: Search HuggingFace for XGBoost variants
- Expected: AUC > 0.85

Python Libraries

• sklearn: scikit-learn models
• xgboost: XGBoost
• lightgbm: LightGBM
• catboost: CatBoost
• optuna: Hyperparameter optimization
• mlflow: Experiment tracking
• huggingface_hub: Model retrieval

Output Format

After training, provide:

1. RAP Retrieval Results - Models found from external sources
2. Model Comparison - All candidates with metrics
3. Best Model - Selected model and configuration
4. Hyperparameters - Optimal parameters found
5. Performance - Test set metrics
6. Model Files - Paths to saved models
7. Recommendations - Deployment notes