Dataset · nvidia
NVIDIA synthetic scientific reasoning dataset
nvidia/OpenScienceReasoning-2
Paired questions and answers with detailed reasoning chains in STEM, law, and economics for LLM SFT fine-tuning.
What's inside
This is a synthetic text dataset containing between 100 thousand and 1 million examples. It includes open-ended questions and multiple-choice tasks augmented with detailed step-by-step reasoning traces. The materials cover exact sciences (STEM), law, economics, and humanities.
What you can do with it
- SFT fine-tuning of language models to enhance logical inference skills
- Preparing models for complex scientific benchmarks such as GPQA-Diamond and MMLU-Pro
- Size
- 100K–1M rows
- Downloads this month
- 39.6K
- Likes
- 58
Models trained on it
- mradermacher/Nemotron-SEA-LION-v4.8-30B-A3B-Base-i1-GGUF 4K
- mradermacher/NeoHorse-1-9B-Math-SFT-i1-GGUF 2.3K
- mradermacher/NeoHorse-1-9B-Math-SFT-GGUF 771
- mradermacher/Qwen3-4B-MegaR3ASONER-v1-i1-GGUFText generation 656
- mradermacher/SciReason-LFM2-2.6B-i1-GGUF 509
- mradermacher/OpenScienceReasoning-Qwen-e10-i1-GGUF 475
- mratsim/MiniMax-M2.5-FP8-INT4-AWQText generation 384
- mradermacher/Qwen3-4B-MegaR3ASONER-v1-GGUFText generation 382
How to load it
from datasets import load_dataset
ds = load_dataset("nvidia/OpenScienceReasoning-2")