Unit 6: Large language models, briefly
About this unit
An LLM predicts a probability for every possible next token; sampling one, appending it and repeating writes a reply, and its billions of parameters come from pretraining on huge text, then fine-tuning.
By the end, you can: Explain an LLM as a next-token predictor run in a loop, trained by pretraining and then fine-tuning.
Modules
4 modulesLearn
- Predicting the next tokenArticle · Tokens, the next-token distribution, the sampling loop and temperature.Not started
- Where the numbers come fromArticle · Billions of parameters, pretraining, fine-tuning with RLHF, and a first look at transformers and attention.Not started
- The whole pictureReview · Prediction, sampling, temperature and the two training stages in one place.Not started