debug-ai-model

What problem does it solve?

This Skill provides a structured, repeatable approach to debugging AI models that fail to converge, produce NaNs, or show poor metrics by guiding you through data, forward pass, loss, gradients, and metrics checks.

Core Features & Use Cases

• Systematic data verification: shapes, dtypes, distributions, and target alignment before model runs.
• Overfit a single batch: verify the model can memorize a single batch and inspect gradients to detect architecture or optimization issues.
• Guided debugging of forward pass, loss, and metrics: ensure correct loss function, proper reduction, and metric implementations across train/val.

Quick Start

Follow the systematic checklist on your failing AI model to identify data issues, overfitting on a single batch, and gradient or loss problems, then implement the fixes.