training-archive-gating
Installation
SKILL.md
Training Archive & Model Gating System (v2.4.1)
Experiment Overview
| Item | Details |
|---|---|
| Date | 2024-12-24 |
| Goal | Create mandatory post-training archival with automatic model classification |
| Environment | alpaca_trading.training package, Colab/local training |
| Status | Success |
Context
Training runs produce summaries with validation metrics, but there was no:
- Structured archival - Summaries were ephemeral JSON files
- Model classification - No clear criteria for deployment readiness
- Overfitting detection - No automatic checkpoint recommendations
- Historical reference - No way to track model improvements over time
The solution: TrainingArchiveManager with automatic gating based on fitness, profit factor, consistency, and drawdown thresholds.