FDE Times: DeepLearning.AI’s Post-training of LLMs — 76 minutes on when to use SFT, DPO or RL
Banghua Zhu’s short DeepLearning.AI course will not teach GPU-cluster ops, but it answers the question FDEs hear most on site: should we fine-tune, and if so with SFT, DPO or online RL? Verifiable rewards (GRPO) are the closest fit to FDE eval work.
FDE Times reviews DeepLearning.AI’s “Post-training of LLMs,” taught by Banghua Zhu (University of Washington, Nvidia, Nexusflow). About 76 minutes of video cover Supervised Fine-Tuning, DPO and online RL (PPO/GRPO), with the goal of choosing the right method rather than running one technique in isolation.
The review’s FDE takeaway is that each method needs a different signal: SFT needs labelled input–output pairs, DPO needs preferred vs rejected answers, and RL needs a scoring function. On a customer site, inventory what data already exists before talking about models—and prefer retrieval when the ask is “know more documents,” not behaviour change.
The GRPO-with-verifiable-rewards section is closest to day-to-day FDE work: writing an automatic reward (valid JSON plus matching waybill, invoice totals that sum) forces the same “what does correct mean?” definition used for evals, even if you never train. Infrastructure at hundred-GPU scale is out of scope; NeMo RL docs are the next step when a project truly needs it.
Source: fdetimes.net