It lowered training costs by quite a bit. To learn from preference data (whats termed as alignment with human values), we used a very large reward model as a proxy for human feedback.
They completely got rid of this, hence also the need to have very large clusters
This has serious implications for spending though. Big companies who would have to train foundation models coz they couldnt directly use meta’s llama, can now just use deepseek.
and directly move to the human/customer alignment phase, which was already significantly cheaper than pretraining (first phase of foundation model training). With their new algorithm, even the later stage does not need huge compute
so they def got rid of a big chunk of compute by not relying on what is called a “reward” model
Unfortunately, that’s not very clear without more. What kind of reward model are they talking about?
This is potentially a 1000x difference in required resources here, assuming you believe their DeepSeek’s quoted figure for spending, so it would have to be an extraordinary change.
From someone in the field
https://github.com/huggingface/open-r1
Unfortunately, that’s not very clear without more. What kind of reward model are they talking about?
This is potentially a 1000x difference in required resources here, assuming you believe their DeepSeek’s quoted figure for spending, so it would have to be an extraordinary change.