Supported by the CSTCloud cloud-native environment, SciHorizon is a platform dedicated to evaluating and applying high-quality data and large-scale artificial intelligence models for research and education. In the era of AI4Science, this platform focuses on two main areas: the scientific capabilities of LLMs and the quality of AI-compatible scientific data. Leveraging CSTCloud's robust computing power and data services, it has built a comprehensive evaluation framework and benchmark system to serve as a valuable tool for AI-driven scientific progress.

To evaluate the performance of Large Language Models (LLMs), SciHorizon establishes a framework consisting of five core indicators: Knowledge, Understanding, Reasoning, Multimodality, and Values. These indicators are broken down into sixteen assessment dimensions, such as Scientific Fact Understanding, Numerical Reasoning, and Adherence to Academic Integrity. Both open-source and closed-source LLMs are evaluated using this framework.

To support scientific data recommendations, SciHorizon presents a generalizable framework for evaluating AI-ready scientific data across four main dimensions: Quality, Fairness, Explainability, and Compliance. These are divided into sub-dimensions. While ensuring high data quality, the framework emphasizes enhancing the semantic richness of the data and its machine-actionable capabilities, and provides application-scenario recommendations.

With the support of CSTCloud, the rankings for large model capability assessments and high-quality scientific data are continuously updated. The core project team has published over 200 papers in top international journals and conferences in the fields of scientific data and artificial intelligence. Additionally, the team has led the development of numerous national and international standards, including IEEE international standards, national standards for scientific data, and industry standards for trusted AI and large models.

For more details, please visit the SciHorizon:https://horizon.scidb.cn/