DeepSeek-R1 had an RLHF step in their post-training pipeline (section 2.3.4 of their technical report[1]).

In addition, the "reasoning-oriented reinforcement learning" step (section 2.3.2) used an approach that is almost identical to RLHF in theory and implementation. The main difference is that they used a rule-based reward system, rather than a model trained on human preference data.

If you want to train a model like DeepSeek-R1, you'll need to know the fundamentals of reinforcement learning on language models, including RLHF.

[1] https://arxiv.org/pdf/2501.12948

Yes but these were steps were not used in R1-zero where its reasoning capabilities were trained.

And as a result R1-zero is way too crude to be used directly, which is a good indication that it remains relevant.

As the other commenter said, R1 required very standard RLHF techniques too. But a fun way to think about it is that reasoning models are going to be bigger and uplift the RLHF boat.

But we need a few years to establish basics before I can write a cumulative RL for LLMs book ;)

This is a GREAT book, if you decide to write it in a rolling fashion you'd have at least one reader from the start :)

This feels like a category mistake. Why would R1 make RLHF obsolete?

You meant to ask "Has r1 made SFT obsolete?" ?