Kuiz berkaitan panduan · Sederhana Tahap
Pemantauan Kualiti Data dalam Kuiz Pengeluaran
Kesan isu skema, kesegaran, ketiadaan dan penyajian kereta api sebelum inferens, sambil membezakan kualiti input daripada prestasi model.
soalan 1 daripada 8
Apakah yang dikesan oleh semakan skema dalam suapan inferens langsung?
Pratonton setiap soalan8 soalan
Jawapan dan penjelasan muncul semasa anda mengambil kuiz.
- Apakah yang dikesan oleh semakan skema dalam suapan inferens langsung?
- Unexpected field names, types or structures
- A guaranteed drop in model accuracy
- A causal change in user behavior
- The model's calibration error
- Apakah yang dikesan oleh monitor kesegaran?
- A change in model weights
- Data that have not been updated within the expected time
- A label mismatch in a training batch
- An unexpected prediction class
- Bagaimanakah nilai kategori yang tidak kelihatan harus dikendalikan?
- Drop the model silently
- Map it randomly to a known category
- Apply an explicit unknown-category policy and monitor its rate
- Assume the value is always invalid
- Ketakpadanan yang manakah merupakan contoh pencongan servis kereta api?
- A difference between two model architectures
- A change in GPU memory usage
- A delayed ground-truth label
- A difference between how features are computed in training and inference
- Mengapakah ambang pengedaran yang terlalu ketat boleh menyebabkan isu pengeluaran?
- It guarantees detection of every defect
- Legitimate seasonal changes may trigger false alarms or blocks
- It prevents all schema changes
- It improves model calibration automatically
- What does passing data-quality checks establish?
- The feature relationships are unchanged
- The model is fair across all groups
- Inputs met the configured checks, not that predictions are accurate
- The data are causally valid
- Where might latency-sensitive data checks run?
- In a dashboard without storing results
- All validation only after labels arrive
- Only in the model-training notebook
- Critical schema checks inline, with heavier analysis asynchronously or in batches
- What should a team do when a production data alert fires?
- Trace the source and determine whether the change is expected before updating rules
- Change the schema immediately to silence the alert
- Retrain automatically without checking data
- Delete the alert history