ML Fundamentals & Model Debugging Drill

Apple MLE loops repeatedly include verbal drills on model choice, evaluation, data processing, LLM concepts, ranking / ads models, GPU OOM, and production metric drops. Requirements Be ready for a rapid-fire MLE round covering: Classification pipeline: data cleaning, feature construction, model choice, training, offline eval, online eval, and monitoring. LLM concepts: Transformer structure, SFT, DPO, prompt engineering, agents, MCP, knowledge distillation, and evaluation. Ads / ranking concepts: binary CTR prediction, DCN v1 vs v2, bidding objectives, CPA design, and A/B testing. Production debugging: online performance drops, GPU OOM, bad embeddings, metric regressions, and data drift. Domain adaptation: audio processing, image / video compression, multimodal embeddings, and model optimization. Inference & deployment: LLM latency and memory bottlenecks (KV cache, quantization, batching), and vision-encoder training basics for multimodal and video roles. Notes The round often behaves like a checklist. Short, precise answers beat long tutorials. For every model or metric, connect it to the data shape and product objective: what is the label, what is delayed, what is biased, what can be measured online, and what failure would page the team. When asked to debug a model, structure the answer as data -> training -> serving -> metric -> user segment. For GPU OOM, mention batch size, activation checkpointing, mixed precision, sequence length, model parallelism, and memory leaks in the training loop. Preparation Build a one-page ML system checklist you can recite: data, labels, model, loss, offline metrics, online metrics, serving, monitoring. Prepare compact explanations for Transformer attention, SFT vs DPO, distillation, embeddings, and prompt evaluation. Practice two debugging narratives: online CTR drops after deploy, and training crashes with GPU OOM. For ads roles, review calibration, delayed labels, auction objectives, and why offline AUC may diverge from business metrics.