Explain LLM training and evaluation
Company: Samsara
Role: Machine Learning Engineer
Category: Machine Learning
Difficulty: medium
Interview Round: Onsite
Explain how transformer-based large language models are pretrained and finetuned. Compare instruction tuning and RLHF, and discuss strategies to reduce hallucinations. How would you evaluate an LLM (e.g., perplexity, task accuracy, human evaluation) and monitor it in production? What inference optimizations (e.g., quantization, KV caching) would you apply?
Quick Answer: This question evaluates an engineer's competency in LLM engineering, including transformer pretraining and finetuning, alignment approaches like instruction tuning and RLHF, strategies for reducing hallucinations, evaluation and production monitoring practices, and inference optimizations.