Which sequence represents the standard order of model training stages?

• A. Fine-tuning, alignment, pre-training
• B. Pre-training, fine-tuning, alignment
• C. Pre-training, alignment, fine-tuning
• D. Alignment, pre-training, fine-tuning

Answer: (B)

Training moves from broad language understanding to task-specific behavior, then to alignment with human preferences. First, pre-training lets the model absorb general patterns, grammar, facts, and reasoning from a large and diverse data mix. Next, fine-tuning specializes the model for particular tasks or domains using focused, labeled data, improving performance in those areas. Finally, alignment uses human feedback to shape outputs toward desired safety and usefulness, often through methods like reinforcement learning from human preferences. This order ensures the model starts with broad capability, then becomes task-ready, and finally behaves responsibly.