5 collocation exercises on the ML lifecycle verbs.
0 / 5 completed
1 / 5
The team will ___ the model on the labelled dataset before evaluating its accuracy.
To train a model is the standard ML collocation for the process of fitting a model to data. Train pairs naturally with model, network, and classifier. Teach, learn, and educate anthropomorphise the model and are not used technically — note that the model learns, but you train it. Engineers say "train the model for 10 epochs," so train the model is the correct collocation.
2 / 5
After training, the data scientist will ___ the model's performance against a held-out test set.
To evaluate performance means to measure how well a model does on unseen data. Evaluate is the precise ML term, behind "evaluation set" and "model evaluation." Judge, rate, and grade are too informal or human-centric. Practitioners "evaluate the model on the test set," so evaluate performance is the correct collocation for measuring model quality.
3 / 5
Rather than training from scratch, the team will ___ a pre-trained model on their domain-specific data.
To fine-tune a model means to continue training a pre-trained model on a narrower, task-specific dataset. Fine-tune is the exact term, as in "fine-tune for a downstream task." Adjust, tweak up, and retouch lack the technical meaning. Note the prepositions: "fine-tune on a dataset" / "fine-tune for a task." So fine-tune the model is the correct collocation.
4 / 5
Once validated, the MLOps team will ___ the model to production behind a serving endpoint.
To deploy a model means to put it into production where it serves predictions. Deploy is the standard term, shared with general software delivery. Release out, launch up, and install are not idiomatic for model serving. Engineers "deploy the model to a serving endpoint," so deploy the model is the correct collocation.
5 / 5
In production, the team must ___ the model for drift as the input data distribution changes over time.
To monitor a model for drift means to continuously track its inputs and performance to detect degradation. Monitor is the precise operations term, behind "model monitoring." Watch over, observe at, and check up are informal or grammatically wrong. MLOps teams "monitor the model for data drift," so monitor for drift is the correct collocation.