Skip to content

Dataset · nvidia

49k complex math problems by NVIDIA for RL

nvidia/AceReason-Math

Cleaned corpus of 49,000 verifiable math problems with answers for training reasoning models via RL.

What's inside

The dataset contains 49,000 diverse English math problems and answers selected from NuminaMath and DeepScaler-Preview. NVIDIA authors strictly filtered out unsuitable examples: removing problems with multiple sub-questions, tests, proofs, images, and long complex answers. Created specifically for training mathematical reasoning using reinforcement learning (RL).

What you can do with it

  • Training LLMs for step-by-step mathematical reasoning using RL methods with verifiable rewards
  • SFT fine-tuning and model evaluation on complex tasks with single-answer output
Tasks
Text generation
Languages
English
Size
10K–100K rows
Downloads this month
4.3K
Likes
56

Models trained on it

How to load it

from datasets import load_dataset

ds = load_dataset("nvidia/AceReason-Math")