Fine-tuning & RL for LLMs: Intro to Post-training
by DeepLearning.AI
AMD's VP of AI walks the whole post-training stack, from the fine-tuning math to the production go/no-go checklist.
Overview
Launched in October 2025 as a DeepLearning.AI-AMD collaboration, this course runs 13 hours 19 minutes across 5 modules and 43 video lessons, with 11 graded assignments and code labs executed on AMD GPUs, though the course is explicit that the methods themselves stay hardware-agnostic. Module 1 places post-training in the LLM lifecycle and works through reasoning and RLAIF safety examples. Module 2 is the technical core: data preparation and tokenization, the fine-tuning loss/gradient/weight-update math split over two lessons, hyperparameter tuning, parameter-efficient fine-tuning with LoRA, reward and preference learning, the RLHF training objective, and the PPO and GRPO algorithms. Module 3 treats evaluation as the north star, covering test sets and metrics, RL test environments and update monitoring, reward hacking, a three-lesson error-analysis sequence that walks errors to causes to fixes, and red teaming against real-world failures. Module 4 covers how much data post-training actually needs, data for fine-tuning versus data for RL, synthetic data pipelines, template engineering and Constitutional AI. Module 5 is production: pipeline design, agents, go/no-go promotion rules for RL runs, the data-feedback flywheel, monitoring and observability, infrastructure and a production-readiness checklist. Instructor Sharon Zhou is VP of AI at AMD; in an O'Reilly interview she frames post-training as what turns a model from raw intelligence into 'usable intelligence and practical intelligence,' and argues RL is now hitting the maturity moment SFT reached years earlier. Access requires a DeepLearning.AI Pro membership at $25/month billed annually or $30 monthly, and completion earns a certificate.
At a Glance
- Topic
- Fine-Tuning
- Level
- Intermediate
- Format
- Course
- Cost
- Paid
- Duration
- ~13 hours 19 minutes across 5 modules and 43 lessons, self-paced
- Provider
- DeepLearning.AI
- Hands-on
- Yes — code/exercises
- Certificate
- Available
What You’ll Learn
- ✓Place post-training in the LLM lifecycle and choose between SFT and RL
- ✓Work through the fine-tuning math: loss, gradients and weight updates, step by step
- ✓Apply parameter-efficient fine-tuning with LoRA and tune the hyperparameters that matter
- ✓Implement reward modeling and preference learning, then run RLHF with PPO and GRPO
- ✓Design evaluation sets and RL test environments, and detect reward hacking early
- ✓Run structured error analysis that traces observed errors to causes to concrete fixes
- ✓Build synthetic data pipelines and template engineering for post-training datasets
- ✓Ship a production post-training pipeline with go/no-go promotion rules and monitoring
Highlights
- •Taught by AMD's VP of AI rather than a generalist course author, and the production and infrastructure modules read like operations experience rather than tutorial filler
- •Unusually long for the DeepLearning.AI catalog: 13+ hours and 43 lessons where most courses in the catalog run one to two hours
- •11 graded labs run on real GPUs, so PPO, GRPO and LoRA are implemented rather than only described
- •Devotes an entire module to evaluation, reward hacking and error analysis, which is where most post-training projects actually fail
- •Covers GRPO alongside PPO, which most 2024-era RLHF material predates entirely
Who It’s For
Best For
- ✓ML engineers moving from prompting into actually training models
- ✓Backend and platform engineers who own an LLM in production and need to customize it
- ✓Data scientists who want RLHF and GRPO grounded in the underlying math
- ✓Technical leads deciding whether to post-train in-house or buy the capability
Prerequisites
- •Strong working Python: the labs are code, not drag-and-drop
- •A basic understanding of how LLMs are trained and what a token is
- •Comfort with gradients and loss functions helps in Module 2, though the math is taught
- •A DeepLearning.AI Pro membership; this course is not on the free tier
FAQ
What is Fine-tuning & RL for LLMs: Intro to Post-training?
A five-module DeepLearning.AI course taught by Sharon Zhou, VP of AI at AMD, covering the whole post-training stack: supervised fine-tuning, LoRA, reward modeling, RLHF, PPO and GRPO, evaluation design, reward hacking, synthetic data pipelines and production rollout. It is for engineers who can already write Python and want to turn a pretrained base model into something a product can ship.
Is Fine-tuning & RL for LLMs: Intro to Post-training free?
Fine-tuning & RL for LLMs: Intro to Post-training is a paid resource.
What level is Fine-tuning & RL for LLMs: Intro to Post-training for?
Fine-tuning & RL for LLMs: Intro to Post-training is aimed at a intermediate audience. Recommended background: Strong working Python: the labs are code, not drag-and-drop, A basic understanding of how LLMs are trained and what a token is, Comfort with gradients and loss functions helps in Module 2, though the math is taught, A DeepLearning.AI Pro membership; this course is not on the free tier.
How long does Fine-tuning & RL for LLMs: Intro to Post-training take?
Expect roughly ~13 hours 19 minutes across 5 modules and 43 lessons, self-paced. Most learners work through it at their own pace.
What will I learn from Fine-tuning & RL for LLMs: Intro to Post-training?
You'll learn: Place post-training in the LLM lifecycle and choose between SFT and RL; Work through the fine-tuning math: loss, gradients and weight updates, step by step; Apply parameter-efficient fine-tuning with LoRA and tune the hyperparameters that matter; Implement reward modeling and preference learning, then run RLHF with PPO and GRPO; Design evaluation sets and RL test environments, and detect reward hacking early; Run structured error analysis that traces observed errors to causes to concrete fixes; Build synthetic data pipelines and template engineering for post-training datasets; Ship a production post-training pipeline with go/no-go promotion rules and monitoring.
Topics
Sources
This page was written from 3 sources, 1 on domains other than deeplearning.ai.