Atlas

Post-training & Alignment

Direct Preference Optimization

Preference alignment without the reinforcement learning.

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. L1IntuitionDPO in one idea: train directly on preference pairs, no reward model or RL loop required.5m
    40 XP
  2. L2MechanicsThe mechanics of the DPO loss: log-probability ratios, the implicit reward, and how the objective is derived conceptually.8m
    70 XP
  3. L3CodeImplement the DPO loss and a full training step end to end, including reference log-prob caching.12m
    110 XP
  4. L4FoundationsThe full derivation from the KL-constrained RLHF objective to the DPO loss, and its gradient's implicit reward-model interpretation.15m
    180 XP

33 stars in the atlas.