AI BriefingOpenR1-Math-220k Dataset Released
Key point
HuggingFace has released OpenR1-Math-220k, a mathematical reasoning dataset aimed at reproducing DeepSeek R1's reasoning capabilities.
Details
HuggingFace's Open R1 Project has announced OpenR1-Math-220k, the first large-scale mathematical reasoning dataset aimed at reconstructing DeepSeek R1's training pipeline and synthetic data.
Key features of the dataset:
- 220k high-quality data points: Based on NuminaMath 1.5, it includes 220,000 math problems and reasoning traces with verified answers.
- Locally generated on H100s: Generated locally using 512 H100 GPUs without relying on APIs. In particular, SGLang was adopted, achieving roughly 2x faster generation speed compared to vLLM.
- Sophisticated filtering:
Math VerifyandLlama3.3-70B-Instructwere used to improve data quality, including cases difficult to verify with rule-based parsers.
Fine-tuning Qwen-7B-Math-Instruct on this dataset achieved performance on par with DeepSeek-Distill-Qwen-7B, demonstrating the dataset's effectiveness.
This summary was generated automatically by AI. Check the original for the author's claims and context. Copyright belongs to the original author.
Our guide explains how the AI works. Report summary errors, attribution issues, or removal requests via Contact.
Related
LLM Mathematical Excellence Driven by Pre-training Data Quality, Not Verifiability