Build harmful-content text classifier
Quick Overview
This question evaluates a candidate's competence in designing an end-to-end machine learning pipeline for binary text classification, covering data understanding and labeling quality, preprocessing, model selection and training, evaluation and thresholding, handling class imbalance and ambiguous labels, and deployment considerations including latency, monitoring, and safety. Commonly asked in the Machine Learning domain, it gauges both practical application skills and conceptual understanding by testing an engineer's ability to balance model performance, evaluation metrics, and operational constraints in real-world NLP and safety-sensitive systems.
You are given a text dataset and asked to build a model that predicts whether a piece of content is harmful (binary classification).
Task
- Propose an end-to-end approach to train and evaluate a classifier (you may assume you can fine-tune a pretrained Transformer).
What to cover
- Data understanding, labeling quality, and preprocessing
- Model choice and training procedure
- Evaluation metrics and thresholding
- Handling class imbalance and ambiguous labels
- Deployment considerations: latency, monitoring, safety/abuse, and model updates