Vol I · Section 2
Training, tuning & inference
Pretraining, fine-tuning, RLHF, and the economics of serving a model.
1Explain pretraining, instruction tuning and RLHF / DPO.2How do you decide between prompting, RAG and fine-tuning?3What is LoRA, and why is it the default way to fine-tune?4How do you evaluate whether a fine-tune actually worked?5Walk through inference latency: TTFT, throughput, batching and the KV cache.6What are the trade-offs of quantizing a model?7Open-weights model or a hosted API — how do you decide?8How do you build a cost model for an LLM feature before launch?9Explain speculative decoding and mixture-of-experts.10How do you upgrade to a new model version without regressions?