Dataset · nvidia
Synthetic scientific questions with reasoning traces
nvidia/OpenScience
Multidomain question-answer pairs with options and reasoning traces for training reasoning LLMs.
What's inside
The dataset is synthetic, covering STEM, law, economics, and humanities. Inside are multiple-choice question-answer pairs with detailed reasoning traces. Designed for SFT or RL to improve accuracy on GPQA-Diamond and MMLU-Pro.
What you can do with it
- SFT LLMs for scientific reasoning
- RL training on multiple-choice questions
- Testing models on GPQA-Diamond and MMLU-Pro
- Size
- 1M–10M rows
- Downloads this month
- 18.9K
- Likes
- 83
Models trained on it
- mradermacher/elf-1-12b-GGUF 737
- mradermacher/elf-1-31b-GGUF 710
- Polygl0t/LilMoo-v0.2Text generation 402
- Polygl0t/LilTii-v0.2Text generation 239
- Polygl0t/Tucano2-0.6B-BaseText generation 224
- mradermacher/sarashina2.2-3B-OpenScience-NoReasoning-300k-GGUF 172
- mradermacher/Tucano2-0.6B-Base-GGUF 158
- mradermacher/LilMoo-v0.2-GGUF 131
How to load it
from datasets import load_dataset
ds = load_dataset("nvidia/OpenScience")