Skip to content
X · @cwolferesearch · X / Twitter

Small detail I caught when re-reading DeepSeek-v4 writeup: they add some ratio of agentic trajectories into mid-training. Interestingly, this is the o…

Small detail I caught when re-reading DeepSeek-v4 writeup: they add some ratio of agentic trajectories into mid-training. Interestingly, this is the only sentence in the paper that mentions the mid-training process used for DeepSeek-V4. But, it seems that blending post-training data into mid-training (rather than just