Table of Contents
What Are Main Training Stages for Modern LLMs Pre-Training Tuning?
Master LLM training’s two core stages—pre-training on vast data, instruction tuning for tasks—with phase breakdowns vs. RLHF myths, for generative AI certification exam prep success. (141 characters)
Question
(Select all that apply) What are the two main stages of the training process for modern LLMs?
A. Pre-training on massive datasets
B. Reinforcement Learning from Human Feedback
C. A linear layer and Softmax function
D. Instruction Tuning
Answer
A. Pre-training on massive datasets
D. Instruction Tuning
Explanation
Modern LLMs undergo a two-stage training paradigm where pre-training on massive, diverse unlabeled datasets (trillions of tokens from web crawls, books, code) first builds general language understanding via next-token prediction or masked modeling, establishing foundational capabilities in grammar, facts, and reasoning across domains. Instruction tuning—also called supervised fine-tuning (SFT)—then adapts this base model using curated instruction-response pairs to enhance following user directives, improving task-specific performance, coherence, and instruction adherence before optional alignment stages like RLHF.
Option B represents an alignment phase post-instruction tuning, not a main training stage. Option C describes output layer components irrelevant to training stages.