Dataset · zwhe99
Complex Verifiable Math Problems for RL
zwhe99/DeepMath-103K
103K complex math problems with verifiable answers, decontaminated, for RL training.
What's inside
Contains a large-scale set of complex math problems with verifiable answers designed for RL. The data is decontaminated; in May 2025, the authors fixed 48 examples where prompts revealed the answer. The language is English, and the size is between 100K and 1M examples.
What you can do with it
- RL fine-tuning of mathematical LLMs
- Evaluating reasoning models on complex tasks
- Training a solution verifier
- Tasks
- Text generation
- Languages
- English
- Size
- 100K–1M rows
- Downloads this month
- 45.6K
- Likes
- 394
Models trained on it
- mradermacher/NeoHorse-1-9B-Math-SFT-i1-GGUF 2.3K
- mradermacher/eubiota-planner-8b-i1-GGUFReinforcement learning 2.2K
- mradermacher/Kvark-Sparkle-1.0-27B-i1-GGUF 2K
- mradermacher/Kvark-Sparkle-1.0-27B-GGUF 1.7K
- mradermacher/Atomight-V2.5-1.7B-C1-i1-GGUF 1K
- liqiangx/InnerVoice-Qwen3-4BText generation 974
- mradermacher/Qwen3-14B-Cobalt2-i1-GGUF 910
- mradermacher/DeepMath-Zero-Math-7B-i1-GGUF 819
How to load it
from datasets import load_dataset
ds = load_dataset("zwhe99/DeepMath-103K")