Hugging Face drops OpenR1-Math-220k: 800k reasoning traces from 512 H100s match DeepSeek's results

Two weeks into the Open R1 project and the community has its first major deliverable: the OpenR1-Math-220k dataset. Hugging Face collaborated with Numina to generate 800,000 reasoning traces from DeepSeek R1 across 400,000 math problems, using 512 H100 GPUs running locally—no API calls required. After an automated filtering pass with Math Verify and a Llama-3.3-70B-Instruct judge, the team landed on a clean set of 220,000 problems with at least one correct chain-of-thought trace. The kicker? Fine-tuning Qwen-7B-Math-Instruct on this dataset matched the performance of the official DeepSeek-Distill-Qwen-7B model. That’s a concrete signal that the distillation pipeline can be replicated independently and at scale.

Generating these traces chewed through a lot of compute, but the engineering choices matter here. The team initially used vLLM for inference, hitting 15 generations per hour per H100. Switching to SGLang nearly doubled throughput to 25 generations per hour per GPU, letting the cluster pump out 300,000 problem solutions daily. They set a 16,000-token generation limit because analysis showed that while 75% of problems solved in under 8,000 tokens, the rest needed every bit of the full context window. Generating two to four solutions per problem also opens the door to rejection sampling and preference optimization methods like DPO—exactly the kind of flexibility that keeps this dataset useful beyond a single fine-tuning run.

The broader context here is about filling a conspicuous gap. DeepSeek published impressive distillation results—their 7B model hit 55.5% on AIME 2024—but never released the 600,000 reasoning traces used to achieve them. Open datasets like OpenThoughts-114k and Bespoke-Stratos-17k had already sprung up to fill the void, but none at this scale with verified correctness. Hugging Face’s contribution proves that an open pipeline can match the proprietary one, and the throughput numbers they’re sharing give a realistic picture of what it costs in GPU hours to pull it off. That kind of transparency is rare and genuinely useful for teams planning their own data generation efforts.

What’s next is the obvious extension: math is a clean, verifiable domain, but code generation poses a harder verification challenge. The team hinted at moving in that direction, and a successful port would make this pipeline template far more valuable. The SGLang speedup alone is worth paying attention to—it’s the kind of infrastructure win that quietly lowers the barrier for everyone else. If the same generation-and-filter workflow can produce competitive code datasets, the Open R1 project won’t just be about reconstructing DeepSeek’s work but surpassing it in practical utility.