Skip to content

Book 04

How the model is trained

First the model practices next-word on a huge amount of text. Later it is shown example replies. Still later, scores and A-versus-B picks change what it prefers.

After this: Name the melting pot, the examples, then the flashlight. · 7 chapters · about 40 minutes

  1. 01First practice: the melting potPretraining5 min
  2. 02What the pot leaves behindFluent — and still sure Sydney is the capital5 min
  3. 03Show it the answer you wantDemonstrations5 min
  4. 04This is a question, not a blogInstruction following5 min
  5. 05A sticky note on a huge bookLoRA5 min
  6. 06One sofa vs a rare platypusScarce cases punch above their count5 min
  7. 07A flashlight, not a report cardEvals5 min

Book 04 examNext: How preferences steer it