'Build a Reasoning Model From Scratch' part 5: scoring answers and self-refinement
A post introducing the tutorial outlines token-probability calculations in PyTorch and a self-refinement loop for generating critiques and revised answers.
TLDR
The shared chapter list progresses from rule-based answer scoring to token probabilities, numerical stability and scoring answers with average log probabilities. It also lists sections on generating critiques and revised answers, implementing a self-refinement loop and examining MATH-500 evaluation results.
