How LLMs Are Trained
Next-token prediction at scale, then instruction tuning and preference training, with the loss, the compute and the limits spelled out.
The one task
An LLM (large language model) is trained on a single objective: given the tokens so far, predict the next token.
- Input: "The cat sat on the"
- Target: " mat"
- Output: a probability for every token in the vocabulary.
That is all. Writing, translating, coding and answering questions all emerge from getting good at this prediction on enough varied text.
1 of 7