CONCEPT Cited by 1 source
Multi-cadence incremental training¶
Definition¶
Multi-cadence incremental training is a model freshness strategy that combines infrequent large-scale retraining passes with frequent lightweight checkpoint updates, balancing freshness against computational cost and catastrophic forgetting.
Netflix GenPage Implementation¶
At Netflix scale, daily retraining of a large transformer from scratch is prohibitively expensive. GenPage uses:
- Periodic broad passes (tunable cadence) — full pretraining + post-training on a wide historical data window
- Daily incremental updates — continue post-training from previous day's checkpoint using latest day's data mixed with sampled historical data
The historical data sampling prevents overfitting to recency while the daily updates capture: - New catalog additions (shows, movies, games) - Shifting user trends and cultural moments - Vocabulary evolution (new entity/row tokens)
New tokens are handled via fallback tokens initialized from type-specific fallback embeddings (Source: sources/2026-06-29-netflix-genpage-generative-homepage-construction).