AISafetyBenchExplorer: Building a Systematic Knowledge Base for AI Safety Benchmarks

AISafetyBenchExplorer is an open-source research tool designed to address the fragmentation issue of AI safety evaluation benchmarks. Through structured directories and multimodal extraction pipelines, it provides standardized metadata management and complexity classification systems for over 180 AI safety benchmarks, making evaluations searchable, comparable, and reproducible. Its core value lies in its dual architecture (manually maintained directory + automated extraction pipeline), which balances accuracy and scalability.

With the evolution of large language model capabilities, AI safety issues have gained attention, but the endless emergence of safety benchmarks has left researchers facing selection difficulties: How to choose benchmarks suitable for specific scenarios? How to compare indicators across different benchmarks? How to balance dataset complexity and coverage? This fragmentation has spurred the demand for systematic knowledge management tools.

The core architecture of AISafetyBenchExplorer includes a manually maintained high-quality benchmark directory (which already includes over 182 benchmarks, each containing 22 standardized fields such as name, task type, evaluation metrics, etc.) and an automated metadata extraction pipeline. Its features include: 1) Controlled vocabulary classification (e.g., safety, jailbreak testing, etc.) for semantic retrieval; 2) Decision tree-based complexity classification (popular, high/medium/low complexity); 3) Standardized recording of evaluation metrics (including LaTeX mathematical definitions, etc.) to support cross-benchmark comparison.

To address the workload issue of manual maintenance, the project has built a multimodal extraction pipeline: using DOI/arXiv ID as the entry point, integrating four academic APIs such as Semantic Scholar to obtain metadata, then using large language models for structured extraction. The pipeline includes data aggregation, PDF parsing, core extraction (instructor framework + OpenAI/Ollama), cross-validation, and other steps. In addition, it provides an AI-assisted extraction process (guided by a main prompt through five phases of work) to enable human-machine collaboration.

1. Researchers: Quickly survey existing benchmarks and filter by scenario (medical AI, finance, etc.) to avoid duplication; 2) Benchmark developers: Refer to metadata standards and use research gap heatmaps to identify gaps; 3) Industry teams: Quantify benchmark maturity through repository activity statistics (star count, maintenance status, etc.) to assist in technology selection.

In engineering: 1700 lines of Python code are modularly organized, Pydantic models ensure type safety, and CLI provides flexible usage. A dual-license strategy (Apache 2.0 for code, CC-BY 4.0 for data and documentation) balances rights and dissemination. In academia: DOI/arXiv integration ensures citation accuracy, and Google Sheets integration lowers the barrier for non-technical users.

The significance of AISafetyBenchExplorer lies in establishing a scalable and maintainable knowledge management framework that helps accumulate collective wisdom, avoid redundant work, and promote the formation of standards. In the face of increasing complexity of safety evaluation brought by breakthroughs in AI model capabilities, such tools provide a structured response method, allowing research to stand on the shoulders of predecessors.