using-training-optimization

What it does

Routes you to the right training optimization specialist based on symptoms and training problems. It emphasizes diagnosing before routing and asks clarifying questions to determine the appropriate sub-skills to apply for issues like model not learning, instability, overfitting, slow training, hyperparameter selection, and experiment management.

How it works

It instructs the agent to diagnose symptoms first and then route to specific skills such as learning-rate-scheduling, gradient-management, loss-functions-and-objectives, data-augmentation-strategies, overfitting-prevention, optimization-algorithms, batch-size-and-memory-tradeoffs, experiment-tracking, and training-loop-architecture. Routing decisions are guided by symptom-specific decision trees and diagnostic questions (e.g., questions about loss behavior, NaN values, optimizer settings, and LR schedules). For ambiguous inputs, it specifies asking one clarifying diagnostic question before routing, and it favors multi-skill, sequential routing when appropriate (e.g., new project setup or convergence issues). It also defines cross-pack considerations (e.g., consulting pytorch-engineering for data loading, profiling, or distributed training) and provides high-level prioritization rules (e.g., diagnose before applying a single technique).

When to use it

Use when there is a training problem and you’re unsure which specific technique to apply. Target scenarios include model not learning, training instability, overfitting, slow training, hyperparameter selection, experiment management, convergence issues, or setting up a new training pipeline. It lists explicit scenarios and when not to use this pack (e.g., PyTorch implementation bugs or production deployment).

What it can touch

• Tools and skills named in the routing paths: learning-rate-scheduling, gradient-management, loss-functions-and-objectives, data-augmentation-strategies, overfitting-prevention, optimization-algorithms, batch-size-and-memory-tradeoffs, experiment-tracking, training-loop-architecture, pytorch-engineering, neural-architectures, llm-specialist, ml-production, deep-rl, etc., as referenced in the skill’s routing guidance.

Caveats

• Contains explicit boundaries about when to avoid using this pack (e.g., CUDA out of memory, architecture selection, deployment concerns).
• Emphasizes diagnosis before routing and clarifying questions before making routing decisions.
• Licensed under MIT in the repository context (as described by the skill metadata).