LLM Training Data (Math)

Product Overview

This product is a Mathematics-domain LLM Training Data Package covering the complete mathematical spectrum from elementary competition math to graduate-level pure mathematics. Contains high-quality math problems, multi-step reasoning processes, proofs, and problem-solving strategies, directly applicable for LLM fine-tuning (SFT), mathematical reasoning reinforcement learning (RLHF / GRPO / Rule-based RL), and preference alignment training.

• Total Data: ~18M+ math QA pairs and reasoning chains
• Token Scale: ~12 billion tokens (including complete proofs/solutions)
• Difficulty Range: AMC 8 → AMC 10/12/AIME → Undergraduate → IMO level → Graduate/Research frontier
• Data Format: JSONL / Parquet (with structured field annotations)
• Quality Standards: Verifiable answer formats, multi-round deduplication, benchmark contamination detection, mathematician spot-checks

---

Covered Mathematical Domains

---

Data Quality & Features

• Complete Multi-Step Reasoning (CoT): Every problem includes full reasoning chains from problem parsing → analysis → strategy selection → step-by-step computation → final conclusion
• Multi-Solution Annotations: ~35% of competition-level problems include 2-4 alternative solution methods (algebraic/geometric/combinatorial)
• Verifiable Answer Formats: Numerical answers and proof conclusions in standardized verifiable formats, supporting Rule-based RL automatic reward signals
• Difficulty Grading 1-10: Level 1-3 (elementary/middle school) → 4-6 (high school/undergrad) → 7-9 (IMO/graduate) → 10 (frontier open problems)
• LaTeX Standardized: All mathematical formulas and symbols in standard LaTeX rendering format
• Benchmark Decontamination: Filtered against GSM8K, MATH, NuminaMath, DeepMath-103K data leakage
• Competition Sources: Includes content from AMC/AIME/IMO/Putnam/CMO competition problems and variants

---

Compatible Evaluation Benchmarks

• MATH Benchmark: 12,500 competition-level math problems across 7 sub-domains (Hendrycks et al.)
• GSM8K: 8,500 grade school math problems testing basic reasoning chains
• NuminaMath: 850K+ human-verified competition-level math problems
• DeepMath-103K: 103K high-difficulty verifiable problems designed for RL
• MMLU-Math / ARC-Math: Math subsets of general evaluations
• MathBench / MathVista: Mathematical visual reasoning and multimodal math evaluations

---

Use Cases

• Math Reasoning Specialist Models: Inject proof & computation capabilities into LLMs (cf. DeepSeek-Math, Qwen2.5-Math training methods)
• Competition-Level Math AI: Train models capable of solving AMC/AIME/IMO-level problems
• EdTech: Intelligent math tutoring systems, automatic solution engines, adaptive problem generation
• Research Computing: Assist proof verification, conjecture exploration, symbolic computation
• Finance & Engineering: Financial pricing models, risk computation, engineering optimization
• RL Reward Training: Verifiable answer formats natively support Rule-based RL / GRPO paradigms

---

Delivery Information

• Data Files: JSONL / Parquet (as agreed), with README and field descriptions
• Delivery Time: Typically 3-7 business days
• Data Updates: Incremental update service available (new competition problems/paper data)
• Compliance: No personal privacy data; for special compliance needs, communicate before ordering