---

name: interpret-curves description: Analyze training logs and explain what's happening. Use when reviewing training runs, understanding loss curves, or diagnosing completed training.

Training Curve Analysis

Analyze existing training logs to understand what happened.

Step 1: Load Logs

Support CSV, JSON, TensorBoard, W&B, or pasted values

Step 2: Plot Overview

Loss curves (train/val), learning rate, gradient norm in 2x2 grid

Step 3: Detect Patterns

• Healthy training: steady decrease, train/val tracking, stable gradients
• Loss plateau: detect onset, duration, percentage of training
• Overfitting: detect onset, final gap
• Instability: coefficient of variation, severity
• Divergence: detect onset, preceding signals

Step 4: Statistical Summary

Overview stats, dynamics assessment, key metrics table, recommendations

Step 5: Compare Runs (if multiple)

Overlay plots, comparison table with final/best loss and timing