DPO from Scratch

Implement log-prob utilities, Bradley–Terry preferences, DPO gradients, IPO variants, and a full evaluation pipeline.

Hard2.3 hours27 steps

Overview

Implement log-prob utilities, Bradley–Terry preferences, DPO gradients, IPO variants, and a full evaluation pipeline.

What you'll learn

Preference learning Alignment Evaluation

FULL WALKTHROUGH

Learn every step on its own page

This project is no longer compressed into a few chapters. Open the dedicated learning workspace for a lesson-by-lesson explanation with concepts, MathJax mathematics, code, tests, mistakes, checkpoints, and persistent navigation.

Open 27-step walkthrough →

Build progress

0 / 27 steps0%

Move the tracker as you finish the original Deep-ML steps. Reaching 100% unlocks the completion action and certificate.

Architecture

Work through the system one dependable layer at a time. Each stage feeds the next and remains independently testable.

01DPO loss→

02Reference policy→

03IPO

∑

Mathematics & visual explanation

Translate the core equations into code, validate intermediate tensors, and compare the implementation with a small numerical reference.

objective(θ) = data_term(θ) + λ · regularization(θ)

The exact objective evolves with each milestone. Keep a notebook of shapes, invariants, and numerical checks.