Open R1: Update #2
Hugging Face
Hugging Face dropped OpenR1-Math-220k, a giant open dataset of R1-style math reasoning traces, built on 512 H100s. It gets a 7B model near DeepSeek's own distilled model, without needing DeepSeek's secret data.
Based on reporting by Hugging Face — read the original for the full story.
Summary, retelling and take written by AI under human oversight; images are AI-generated illustrations. How we work · Report an error
Two weeks into the Open R1 project, the Hugging Face team has moved past theorizing about DeepSeek R1's training pipeline and actually shipped something usable: OpenR1-Math-220k, a 220,000-problem dataset of verified math reasoning traces. The pitch is simple. DeepSeek never released the 600k reasoning traces it used to distill R1's abilities into smaller Qwen and Llama models, so the community has been stitching together its own version, piece by piece.
The production numbers here are the part that stands out. Running DeepSeek R1 locally on 512 H100 GPUs, the team generated 800,000 reasoning traces across 400,000 problems pulled from NuminaMath 1.5. Early on they leaned on vLLM and got 15 generations per hour per GPU; switching to SGLang nearly doubled that to 25, letting them crank out 300,000 solutions a day. That's the kind of throughput that used to require API access to a frontier lab, not a university-scale cluster and some clever engineering.
Filtering turned out to be almost as important as generation. Math Verify, a rules-based checker, confirmed correct answers for 55% of problems, but plenty of NuminaMath's ground-truth answers were malformed or blank, so the team brought in Llama-3.3-70B-Instruct as a judge to rescue another 28,000 problems that automated verification had wrongly tossed. They also tried using a 72B reward model to pick the best of multiple correct generations per problem, which sounds like it should help — except it didn't. Training ablations showed picking a random correct answer worked just as well as the RM-ranked pick, a small but telling reminder that fancier pipelines don't automatically buy better models.
The payoff: fine-tuning Qwen2.5-Math-7B-Instruct on the 94k
My take — AI-written commentary, not fact-checked reporting
What jumps out to me isn't the 220k number, it's that a random correct answer beat their reward-model-ranked pick — that's the kind of humbling result labs rarely publish, and it's exactly why open reproduction work matters more than another closed benchmark flex. The AIME 2025 leakage finding is the real gut-punch here: if competition problems are already sitting on forums, half of what we call reasoning benchmarks might just be memorization with extra steps.
Read more about this at: Hugging Face