📊 NEXUS Training Metrics
ML Training Analytics & Convergence Analysis
🎯 Training Progress Summary
📝 Key Insights
- • Loading training data...
✅ Pre-event training with 30 episodes demonstrates model's learning capability. Continue training on-site to improve beyond baseline metrics.
Training Progress Overview
Total reward improvement over training episodes. Shows learning trajectory and convergence toward optimal performance.
Reward Dimensions Breakdown
Each reward dimension's evolution over training. Identify which aspects are improving vs plateauing.
Curriculum Learning Progression
DifficultyAdapter effectiveness. Shows distribution of difficulty levels and success rates per level.
Difficulty Distribution
Convergence Analysis
Analyze learning stability, convergence rate, and trend direction.
Before/After Training Comparison
Observable behavior change: untrained vs trained agent performance.
| Metric | Untrained (Baseline) | Trained (After Training) | Improvement |
|---|---|---|---|
| Reward (INC003) | 0.28 | 0.28 | 0.00 (0%) |
| Avg Steps to Resolve | 45 | 45 | 0 steps |
| Coalition Accuracy | 0% | 0% | 0% |
| MTTR Score | 0.15 | 0.15 | +0.00 |
| Diagnosis Accuracy | 25% | 25% | 0% |
| Coordination Quality | 20% | 20% | 0% |
📁 Colab training exports
Interactive charts from learning_curve.json plus Colab-generated PNGs.
Regenerate web copies with python scripts/sync_training_runs_web.py after adding folders under NEXUS_GRPO_backups/.
This tab is separate from live Space /training-metrics (runtime aggregate).