中文科學知識 MCQA 問答數據

# OpenScience-Chinese A Chinese multiple-choice science QA dataset with chain-of-thought reasoning, derived from [nvidia/OpenScience](https://huggingface.co/datasets/nvidia/OpenScience) through translation and rejection sampling. * Key Features: * Scale: 50,000 high-quality instances. * Reasoning: Built-in Chain-of-Thought (`<think>` tags) for interpretable AI. * Reliability: Rejection sampling ensures 100% alignment with ground truth. ## Data Source The questions originate from [nvidia/OpenScience](https://huggingface.co/datasets/nvidia/OpenScience), a large-scale science QA dataset covering diverse domains including physics, chemistry, biology, medicine, computer science, and engineering. The original English questions were translated into Chinese and polished for fluency. ## Construction Pipeline ``` nvidia/OpenScience (English MCQA) │ ▼ Sampling & Translation & Polishing │ Translated to Traditional Chinese with terminology review │ ▼ Rejection Sampling (3 rejections per prompt) │ temperature=0.6, top_p=0.95, max_tokens=16384 │ ▼ Answer Extraction & Verification │ Extract model's chosen answer → compare with ground truth │ Keep only correct responses │ ▼ Sampling & Quality Filtering │ ▼ Final Dataset (50,000 examples) ``` ### Models Used | Model | Samples | Accuracy (num_rejections=3) | |-------|---------|-------| | Kyara-3-Flash | 35,000 | 88.32% | | Luoyang-48B-A3B | 15,000 | 89.48% | ### Rejection Sampling Details - Each question was sampled 3 times (num_rejections=3) with `temperature=0.6` and `top_p=0.95` - Model-extracted answer choices were compared against the ground truth from OpenScience. - Only responses where the model's final answer matched the ground truth were retained. ## Format Each example follows a standard `messages` format with reasoning in `<think>` tags: ```json { "messages": [ { "role": "user", "content": "以下哪一個古代帝國以其在數學與天文學發展上的重要貢獻而聞名，但卻經常被更為人熟知的文明所掩蓋？\n\nA: 孔雀王朝 (Maurya Empire)\nB: 印加帝國 (Inca Empire)\n..." }, { "role": "assistant", "content": "<think>\n這個問題是問哪個古代帝國在數學和天文學上有重要貢獻，但經常被更知名的文明掩蓋。選項有A到J，我需要仔細想想每個選項的情況。\n\n首先，題目中的選項包括孔雀王朝、印加帝國、阿克蘇姆、薩珊王朝、塞伐那王朝、貴霜帝國、阿拔斯王朝、羅馬帝國、笈多王朝、拜占庭帝國。我需要回憶每個帝國在數學和天文學方面的貢獻，然後看哪個可能被其他更出名的文明所掩蓋。...</think>\n在眾多古代帝國中，**笈多王朝（Gupta Empire）**（選項 I）以其在數學與天文學上的重要貢獻而聞名，卻常被更為人熟知的文明（如古希臘、羅馬或阿拉伯帝國）所掩蓋。..." } ], "metadata": { "model": "Luoyang-48B-A3B", "temperature": 0.6, "top_p": 0.95, "num_rejections": 3, "max_tokens": 16384 } } ``` ## Usage ```python from datasets import load_dataset ds = load_dataset("zake7749/OpenScience-Chinese") print(ds["train"][0]["messages"]) ``` ## Limitations - The dataset is focused on Traditional Chinese. - All responses have been verified for answer correctness, but the reasoning paths are not individually verified for logical soundness. - Domain coverage follows the distribution of the source OpenScience dataset, which is skewed toward STEM subjects. ## Citation If you use this dataset, you may cite the dataset as: ```bibtex @misc{yang_openscience_chinese_2026, title={OpenScience-Chinese}, author={Yang, Kai-Chou}, year={2026}, publisher={Hugging Face}, url={https://huggingface.co/datasets/zake7749/OpenScience-Chinese} } ``` And please cite the original OpenScience dataset: ```bibtex @misc{nvidia_openscience, title={OpenScience}, author={NVIDIA}, year={2025}, publisher={Hugging Face}, url={https://huggingface.co/datasets/nvidia/OpenScience} } ```