LLMs from Scratch – Practical Engineering from Base Model to PPO RLHF

The course is an exceptional resource for anyone seeking a deep, practical understanding of LLMs. Its primary strength lies in its comprehensive, end-to-end approach, covering everything from the fundamental transformer architecture to advanced alignment techniques like RLHF with PPO. The instructor, Vivek Kalyanarangan, demonstrates a clear and methodical teaching style, breaking down complex concepts into manageable, code-driven segments. The use of pure PyTorch ensures that learners gain a genuine understanding of the underlying mechanics, rather than relying on high-level abstractions. The course is well-structured, with each part building logically on the previous, and the provided GitHub repository contains modular, well-commented code that facilitates hands-on learning. The theoretical explanations are concise but sufficient, providing the necessary context without overwhelming the learner. The course also incorporates modern techniques such as rotary embeddings, RMSNorm, SwiGLU, and mixture-of-experts, ensuring that learners are exposed to current best practices. The alignment phase, including reward modeling and PPO, is particularly valuable, as it addresses the critical aspect of making LLMs helpful and safe. The main limitation is the lack of formal citations to academic papers or external sources, which could enhance the credibility and allow learners to explore the underlying research. However, the content is accurate and aligns with established knowledge in the field. The title accurately reflects the content, and the course delivers on its promise of building an LLM from scratch. The public comments are overwhelmingly positive, with viewers expressing gratitude and enthusiasm, and many indicating they plan to follow along and build their own models. The course’s practical, hands-on nature is highly praised, and it is considered a valuable contribution to the learning community. Overall, this is an excellent, high-quality resource that deserves a top rating.

The title accurately reflects the content: a comprehensive, from-scratch implementation of LLMs, covering base model training to PPO-based RLHF.

9/10

The course is a hands-on tutorial with code available on GitHub, covering the full LLM pipeline from transformer architecture to RLHF. The content is technically accurate, well-structured, and based on established research (e.g., Attention Is All You Need, PPO). The instructor demonstrates deep expertise and provides practical implementation details. The main limitation is the lack of formal citations to sources, but the code and explanations are reproducible.

• 00:00Part 0 - Introduction
• 05:43Part 1 - Core Transformer Architecture
• 40:24Part 2 — Training a Tiny LLM
• 1:30:27Part 3 — Modernizing the Architecture
• 2:33:53Part 4 — Scaling Up
• 3:17:22Part 5 — Mixture-of-Experts (MoE)
• 3:44:19Part 6 — Supervised Fine-Tuning (SFT)
• 4:23:44Part 7 — Reward Modeling
• 4:59:55Part 8 — RLHF with PPO

• GitHub repository for the course code — All code modules and orchestrator scripts for each part of the course.
• freeCodeCamp news — Platform hosting the course and related articles.
• freeCodeCamp main site — General learning platform.

• Attention Is All You Need — The transformer architecture is based on this paper, which the course implements from scratch.
• Proximal Policy Optimization Algorithms — The PPO algorithm used in the RLHF section is derived from this paper.

• scrimba.com