Yes, but in this case the RL stage doesn't actually need a huge amount of data. Most of the data is ingested in the pre-training phase, which in this case was d

Yes, but in this case the RL stage doesn't actually need a huge amount of data. Most of the data is ingested in the pre-training phase, which in this case was done by Alibaba. There was a recent paper [1] that showed that for one particular model it was possible to get very good reasoning performance by finetuning on under a thousand examples of reasoning. [1] https://arxiv.org/abs/2502.03387 https://arxiv.org/abs/2502.03387