GooseReason-0.7M is a large-scale reinforcement learning with verifiable rewards (RLVR) dataset released by NVIDIA in 2026, with the associated research paper available at Golden Goose: A Simple Trick to Synthesize Unlimited RLVR Tasks from Unverifiable Internet Text. It aims to address the bottleneck of scarce verifiable training data by synthesizing an unlimited number of RLVR tasks from rich reasoning-oriented internet text through the Golden Goose pipeline.

The dataset contains over 700,000 tasks spanning mathematics, programming, and general science, with a data scale of 0.7M level. It is primarily used to train the GooseReason-4B-Instruct model, which achieves new state-of-the-art results among 4B-level models across 15 benchmarks covering mathematics, programming, STEM reasoning, instruction following, and logic puzzles. This dataset is intended for research and development purposes only.

Dataset Composition

• Mathematics (Math): Contains 235,836 samples sourced from AoPS-Instruct (approximately 600,000 QA pairs from the Art of Problem Solving forum, primarily olympiad-level mathematics problems).
• Programming (Code): Contains 281,793 samples sourced from rStar-Coder (approximately 418,000 programming problems from competitive programming platforms, using the synthetic_sft split, containing only problems and teacher model solutions, without test cases).
• Science (STEM): Contains 155,496 samples sourced from MegaScience (approximately 650,000 QA pairs from about 12,000 university-level science textbooks, covering physics, biology, chemistry, medicine, computer science, mathematics, and economics).

Citation

@article{lu2026goldengoose,
  title={Golden Goose: A Simple Trick to Synthesize Unlimited RLVR Tasks from Unverifiable Internet Text},
  author={Lu, Ximing and Acuna, David and Jung, Jaehun and Hu, Jian and Zhang, Di and Diao, Shizhe and Zou, Yunheng and Zhang, Shaokun and Cui, Brandon and Liu, Mingjie and Kim, Hyunwoo and Ammanabrolu, Prithviraj and Kautz, Jan and Dong, Yi and Choi, Yejin},
  journal={arXiv preprint arXiv:2601.22975},
  year={2026}
}