IFM opens K2 Horizon's pre-training data and post-training data-generation pipelines

IFM says it has opened K2 Horizon's pre-training datasets and post-training data-generation code and tools for researchers. It lists 13 trillion pre-training tokens across five datasets and 24 subsets, plus over 10 trillion synthetic tokens spanning mathematics, reasoning, coding, planning and problem-solving. IFM says strict verification can discard useful training signal, and its chosen data mix lost in smaller experiments.