ml-reviewer_skill

Overview

This skill reviews machine learning and deep learning code and configurations for PyTorch, TensorFlow/Keras, and scikit-learn projects. It identifies model architecture issues, training loop mistakes, data pipeline problems, GPU/performance pitfalls, and MLOps gaps. The goal is actionable recommendations to improve correctness, performance, and reproducibility.

How this skill works

The reviewer scans project files and common manifests for ML frameworks, model files, and training scripts. It inspects model definitions, training/validation loops, data preprocessing, DataLoader/dataset usage, GPU transfers, and experiment/config management. For each finding it reports severity, file/line hints, and concrete fixes (code suggestions or tooling).

When to use it

• Code reviews for ML/DL projects using PyTorch, TensorFlow/Keras, or scikit-learn
• When validating model training, inference correctness, or GPU performance
• Before deploying models to production or preparing reproducible experiments
• When adding experiment tracking, config management, or model versioning
• To audit data pipelines for leakage, augmentation, and memory usage

Best practices

• Run inference with model.eval() / tf.function and wrap with no_grad or proper tracing
• Prevent data leakage: fit scalers on train only, use validation split and CV
• Optimize GPU usage: move tensors to device, batch GPU transfers, use mixed precision
• Manage experiments: set seeds, track runs (wandb/mlflow), and externalize hyperparameters
• Set DataLoader num_workers, use tf.data or generators for large datasets, and add augmentation

Example use cases

• Review a PyTorch repo for missing model.eval(), in-place ops, and gradient clipping
• Audit a Keras training loop for missing @tf.function, eager-mode overhead, or OOM risks
• Detect scikit-learn data leakage from fit_transform on test data and missing cross-validation
• Improve GPU throughput by flagging CPU tensor ops, frequent CPU-GPU transfers, or lack of accumulation
• Assess MLOps maturity: missing experiment tracking, hardcoded hyperparameters, or no model versioning