Intermediate
Open
Pro
Use Tracked Runs to Choose Among a Hyperparameter Sweep
A team ran a sweep of 5 configurations for a gradient-boosted model, logged as tracked runs:
| Run | learning_rate | max_depth | val_auc | val_logloss | train_time_min |
|---|---|---|---|---|---|
| A | 0.30 | 4 | 0.831 | 0.412 | 6 |
| B | 0.10 | 6 | 0.857 | 0.379 | 11 |
| C | 0.10 | 8 | 0.862 | 0.371 | 18 |
| D | 0.03 | 8 | 0.849 | 0.390 | 22 |
| E | 0.10 | 10 | 0.859 | 0.376 | 27 |
This model retrains automatically every night, and training compute is billed per minute.
- Which run would you register as the candidate for promotion, and why — considering both quality and the operational cost of nightly retraining?
- What additional information, beyond what's in this table, would you want to check in each run's full tracked record before finalizing that choice?
- Explain, referencing this specific table, why having this comparison available at all depended on these being tracked runs rather than five one-off notebook executions with printed results.
Share this question