Training Debug & Performance Tuning
Debug and optimize model training with systematic loss curve analysis
What You Can Do
You can systematically diagnose why your models aren't converging, identify which hyperparameters matter most for your training dynamics, and generate prioritized optimization strategies. By analyzing loss curves, gradient flow patterns, and validation metrics, you'll accelerate convergence, reduce wasted compute, and achieve better generalization—without trial-and-error tuning.
Features
decode divergence, saturation, oscillation, and overfitting signals to understand what's happening during training
rank which parameters (learning rate, batch size, regularization, warmup) most affect your model's convergence
detect vanishing/exploding gradients, activation saturation, and dead neurons that stall learning
receive prioritized, actionable recommendations to reach target performance faster
systematically distinguish overfitting from underfitting and quantify generalization risk
benchmark warmup strategies, decay schedules, and adaptive methods for your architecture
evaluate computational cost vs. model quality for different configurations
systemize diagnosis of NaN losses, divergence, stalled progress, and instability
Example Output
Example 1: Loss Curve Analysis
Observations:
✓ Training loss decreases smoothly → healthy gradient flow
✗ Validation loss plateaus after epoch 15 → overfitting begins
✗ Loss spike at epoch 8 → likely learning rate too high or batch effect
Root Cause: Validation overfitting, training LR still acceptable
Recommendations:
1. Add L2 regularization (start λ=0.001)
2. Implement early stopping with patience=5
3. Increase dropout to 0.5 in dense layers
Example 2: Hyperparameter Sensitivity Ranking
Impact on Final Validation Accuracy:
1. Learning Rate (±15% variance) — CRITICAL
2. Warmup Steps (±8% variance) — HIGH
3. Batch Size (±5% variance) — MEDIUM
Top 3 Actions:
- Reduce LR from 1e-3 to 5e-4 (steeper convergence)
- Extend warmup from 1000 to 3000 steps (smoother early training)
- Increase batch size 32→64 (30% faster with better gradient estimates)
Example 3: Convergence Acceleration Plan
Current: 85% accuracy in 50 epochs → Target: 87% in 30 epochs
Proposed Changes (estimated 35% speedup):
1. Scheduler: Linear warmup (3k steps) → cosine decay
2. Regularization: L2(0.0001), increase dropout 0.3→0.4
3. Batch: 32→64 with augmentation on 80% of data
4. Expected: reach 85% by epoch 30, potential 87% by epoch 40
What's Included
- SKILL.md: Complete training diagnostics workflow with decision trees for systematic analysis
- Loss Curve Analyzer checklist: Step-by-step guide to interpret patterns and identify failure modes
- Hyperparameter sensitivity template: Framework for ranking parameters by impact on convergence
- Training dynamics worksheet: Prompts to structure your loss logs, metrics, and observations
- Convergence acceleration playbook: Prioritized strategies for your specific bottleneck
- Debugging flowchart: Decision tree to diagnose NaN, divergence, or stalled training
- Learning rate tuning guide: Warmup and decay schedule recommendations by architecture
Who It's For
- ML Engineers building and deploying production models who need faster iteration cycles
- Research Scientists tuning deep learning models and investigating convergence issues
- ML Ops Engineers optimizing training pipelines and reducing computational cost
- Data Scientists training custom models on limited budgets or hardware
- Computer Vision/NLP Specialists debugging architecture-specific training dynamics
Best For
- Debugging training failures (NaN loss, divergence, stalled progress, oscillation)
- Hyperparameter optimization without expensive grid search
- Analyzing and accelerating convergence speed
- Reducing wasted compute by identifying ineffective configurations early
- Improving model generalization and reducing overfitting
You might also like
This skill enables you to automatically audit your websites against compliance standards and detect deviations from expected baselines. You can track regulatory requirements, identify policy violations, and generate compliance reports with actionable remediation steps. Monitor multiple sites simultaneously and maintain detailed audit trails for compliance documentation.
Analyze chemical processes systematically to identify hazards, assess risks, and generate safety recommendations. You can perform HAZOP analyses, evaluate compliance with industry standards, conduct root-cause analysis of incidents, and develop emergency response procedures. The skill guides you through structured safety reviews that reduce the likelihood of accidents and regulatory violations.
You'll develop systematic approaches to injectable formulation design, from API selection through sterilization strategy. Claude helps you troubleshoot failed batches by analyzing root causes, recommends regulatory pathways (505(b)(2), ANDA, NDA), and provides science-backed solutions for stability, compatibility, and manufacturability challenges.
You can conduct comprehensive hazard analyses for chemical processes using industry-standard methodologies like HAZOP and LOPA. The skill helps you assess risks quantitatively, identify control gaps, design engineered safeguards, and generate formal documentation for regulatory compliance and process safety management.
Claude analyzes your code diffs and generates standardized commit messages that follow the Conventional Commits specification. The skill automatically determines the correct commit type, scope, and description based on the changes you've made, ensuring your messages are parseable by automation tools while remaining human-readable for code reviewers.
You can transform raw text into structured, labeled datasets for machine learning, analysis, and research. This skill performs named entity recognition, sentiment classification, part-of-speech tagging, and dependency parsing—generating consistent, validated annotations at scale. Use it to prepare corpora, extract entities, classify documents, or perform linguistic analysis without manual annotation.
This skill automates the creation of production-ready deployment validation checklists, infrastructure-as-code templates, and incident response runbooks tailored to your ML stack. You get comprehensive pre-deployment checks covering model validation, data pipeline integrity, infrastructure readiness, and monitoring setup—all customized for your specific models and cloud provider. The skill generates executable runbooks that teams can follow during incidents, including rollback procedures, failover strategies, and diagnostic commands.
Rapidly analyze firmware logs and diagnose hardware issues that cause device failures, connectivity problems, and performance degradation. You'll identify root causes from stack traces, crash dumps, and sensor data, then generate specific optimization recommendations. This skill transforms raw device logs into actionable debugging plans that reduce time-to-resolution from hours to minutes.