Reinforcement Learning for Large Language Models Making the Hard Math Easy — Companion Repository
Book: [https://github.com/PacktPublishing/Reinforcement-Learning-for-LLMs ]Author: [Arun Shankar & Michael Chertushkin]
Click any badge to open a notebook in Google Colab — no installation needed.
#
Chapter
Colab
1
Essential Math Toolkit
2
Why LLMs Need RL: The Alignment Gap
3
RL Fundamentals: The Complete Picture
4
Setting Up Your Free Environment
#
Chapter
Colab
5
Supervised Fine-Tuning & The Cold Start
6
RLHF: The Three-Step Dance
7
Direct Preference Optimization
8
Online DPO & Iterative Alignment
9
Reward Modeling & The Critic
10
Modern RL Algorithms: GRPO, RLOO, KTO
Part 3 — Advanced Techniques
#
Chapter
Colab
11
Verifiers & Outcome Rewards
12a
Reasoning with GRPO — Concepts
12b
Reasoning with GRPO — The DeepSeek Recipe
13
Test-Time Compute: Scaling at Inference
14
Self-Play & Constitutional AI
15
Multi-Objective RL & Agentic Systems
16
Domain-Specific RL: Code, Math, Tools
Part 4 — Production & Recipes
#
Chapter
Colab
17
Recipe: Chatbot
18
Recipe: Reasoner
19
Recipe: Agent
rl-made-easy-code/
├── notebooks/
│ ├── part1_foundations/ # Chapters 1–4
│ ├── part2_core/ # Chapters 5–10
│ ├── part3_advanced/ # Chapters 11–16
│ └── part4_recipes/ # Chapters 17–19
├── utils/
│ ├── data.py # Dataset loaders
│ ├── eval.py # Win rate, KL divergence helpers
│ └── viz.py # Training curve plots
├── data/samples/ # Toy datasets — notebooks run offline
│ ├── preferences.jsonl
│ └── prompts.jsonl
└── requirements.txt # Installed automatically by each notebook
All notebooks target the free Colab T4 GPU . Runtime → Change runtime type → T4 GPU.
Part
Estimated runtime on T4
Part 1 — Foundations
< 5 min (CPU)
Part 2 — Core methods
10–30 min
Part 3 — Advanced
20–60 min
Part 4 — Recipes
30–90 min
Code: Apache 2.0 · Text & figures: © [Arun Shankar & Michael Chertushkin], All Rights Reserved
