Build a Reasoning Model (From Scratch)

About this event

We are working through Build a Reasoning Model (From Scratch) by Sebastian Raschka. During each meetup, we will discuss the key concepts from the chapter or pages under review. In this session we'll discuss the Chapter 6. Training Reasoning Models with Reinforcement Learning. Raschka provides a hands-on guide to understanding and building modern reasoning models. Starting with a conventional pretrained LLM, the book explores evaluation, inference-time reasoning techniques, reinforcement learning with verifiable rewards, and knowledge distillation. Please note that the session is not recorded, and participants are responsible for obtaining their own copy of the book. Buy the book (affiliate links): Amazon UK (https://amzn.to/4qecrce) Amazon France (https://link.amazon/B00LK2CWI) Book overview: In Build a Reasoning Model (From Scratch), bestselling author Sebastian Raschka explains how modern reasoning-oriented language models work by implementing their core techniques step by step. You will begin with a conventional pretrained LLM and learn how to generate and evaluate its responses. You will then explore inference-time techniques such as chain-of-thought prompting, sampling, self-consistency, response scoring, Best-of-N, and self-refinement. Later chapters introduce training-based approaches, including reinforcement learning with verifiable rewards, GRPO, format rewards, and distillation from stronger reasoning models into smaller ones. Build a Reasoning Model (From Scratch) teaches you how to: • Implement core LLM reasoning techniques from scratch • Generate, score, and evaluate model responses • Build verifier-based evaluation systems • Improve reasoning using self-consistency, Best-of-N, and self-refinement • Apply reinforcement learning with verifiable rewards • Understand and implement GRPO-based training • Distill the reasoning capabilities of stronger models into smaller ones • Evaluate the accuracy, cost, and latency trade-offs of different reasoning techniques ***