Atlas

Post-training & Alignment

Reward Models

A model that scores answers so another model can chase the score.

You can jump straight in, but this star assumes RLHF. Not recommended as a first stop.

Four layers of depth

Each layer ends with a quiz. Finish layer 4 and you own this concept.

  1. L1IntuitionWhat a reward model is: a scorer trained to imitate human preference, not to answer questions.5m
    40 XP
  2. L2MechanicsHow preference data becomes a training signal: the Bradley-Terry model and the ranking loss.8m
    70 XP
  3. L3CodeImplement and train a Bradley-Terry reward model head-to-tail in PyTorch.12m
    110 XP
  4. L4FoundationsThe statistical model behind reward-model training and the effect of miscalibration on downstream RL.15m
    180 XP

33 stars in the atlas.