July 2025 Summaries

SYNTHETIC-2 is an extensive open dataset comprising four million verified reasoning traces aimed at tackling complex reinforcement learning (RL) tasks, collaboratively generated with the help of over 1,250 GPUs worldwide using pipeline-parallel distributed inference. This dataset incorporates a wide range of challenging reasoning tasks, including math, coding, and more unconventional tasks like puzzles, to enhance instruction-following capabilities. It includes novel verifiable tasks such as Code Output Prediction v2 and Pydantic Adherence, with tasks categorized into two subsets for high-quality supervised fine-tuning (SFT) and RL, ensuring both quality and verification through models like DeepSeek-R1-0528. The infrastructure optimizes global orchestration of GPU nodes, ensuring transparency and accountability, while employing TOPLOC v2 for verifiable distributed computations. The dataset, available on Hugging Face, serves as a foundation for future distributed RL runs, with plans to expand the RL environment's ecosystem and enhance the capabilities of state-of-the-art reasoning models through global compute collaborations.