How do you decide when to retrain a model, and how do you design a retraining pipeline that's reliable enough to run automatically?
formulate your answer, then —
You mentioned training from scratch vs. fine-tuning. How do you choose, and what can go wrong with fine-tuning on recent data?
formulate your answer, then —
tldr
Retrain on a schedule as a baseline; add trigger-based retraining for models sensitive to sudden shifts. A retraining pipeline needs data validation, an evaluation gate against the production model, and a rollback path before you can safely automate promotion. Fine-tuning is faster but risks catastrophic forgetting — use a replay buffer of historical data or retrain from scratch when the distribution has shifted significantly.
follow-up
- How would you design a retraining pipeline for a model where ground-truth labels arrive with a 7-day delay?
- What's the difference between online learning (updating on each example) and periodic retraining, and when would you use each?
- How do you prevent a retraining pipeline from silently failing — e.g., training on stale data due to a broken data pipeline?