Multilingual Math Dataset for RLVR

🔑 Enhanced Key Takeaways

• •mAceReason-Math comprises over 140k high-quality translations of challenging math problems across 14 languages, with full train splits containing 10,270–12,245 samples per language.[1][2][3]
• •Dataset derived from a filtered subset of AceReason-Math (Chen et al., 2025), originally curated for GRPO training, with rigorous cleaning distinguishing salvageable surface issues from critical exclusions.[1]
• •Features a fully parallel train split of 7,620 samples per language and a human-validated test set of 190 samples, produced via iterative LLM translation refined by native-speaker checks.[1]

🛠️ Technical Deep Dive

• •Sourced from AceReason-Math corpus for RLVR/GRPO training; base data undergoes multi-stage cleaning pipeline addressing surface-level (fixed and translated) and critical issues (excluded).[1]
• •Translation process: iterative LLM-based with targeted cleanup and native-speaker validation to ensure linguistic fidelity and mathematical accuracy across 14 languages.[1][2]
• •Splits include balanced parallel train (7,620 samples/language), full train (10k+ per language, non-parallel), and human-validated test set (190 samples).[1][3]

🔮 Future ImplicationsAI analysis grounded in cited sources

⏳ Timeline