A foundational model is a broadly pretrained language model whose learned knowledge and capabilities provide the starting point for later adaptation. Pretraining produces broad potential, but the raw model operates as a next-token predictor: it continues text rather than reliably acting as a conversational assistant. Its eventual quality therefore depends both on what pretraining made possible and on how later stages elicit that potential. section 1.4 section 4.0
The practical motivation is to reuse one model across tasks instead of building a separate system for each one. Prompting and in-context examples can steer a pretrained model toward a task, while instruction fine-tuning more directly trains a common instruction–response interface. This makes task generalization more reliable and gives subsequent alignment stages a model that already has the expected shape of an answer. section 4.0 section 1.3
A typical adaptation pipeline separates capabilities from behavior:
- Begin with broad pretraining, which creates knowledge and capabilities but leaves interaction governed by next-token continuation. section 1.4 section 4.0
- Instruction-tune on high-quality question–answer examples using the same next-token loss, shifting outputs toward an assistant persona. section 1.3
- Train a reward model on preference relations to score answer quality, then use reinforcement learning to improve helpfulness. section 1.3
- Iterate on post-training choices such as end-of-pretraining data, instruction fine-tuning, RL with verifiable rewards, and preference tuning. section 1.4
Continuation versus instruction following
Given “What is the boiling point of water?”, a base language model may continue into related chemistry questions rather than answer. After instruction fine-tuning, the same prompt is answered directly with 100C (212F) at standard atmospheric pressure. The contrast shows that adaptation changes the interface and response behavior, not merely the model’s stored knowledge. section 4.0
Post-training is not a substitute for the base model’s underlying potential. It is a process for making that potential accessible in useful interaction, analogous to improving a car through changes after its chassis and engine are fixed. In one cited open mixture-of-experts case, retaining most pretraining while changing post-training raised an evaluation average from 35 to 48, illustrating how strongly adaptation can affect measured performance. section 1.4
