🧜‍♀️ Merbench - LLM Evaluation

Getting LLMs to consistently nail the mermaid diagram syntax can be... an adventure.

Merbench evaluates an LLM's ability to autonomously write and debug Mermaid syntax. The agent can access an MCP server that validates its code and provides error feedback, guiding it towards a correct solution.

Each model is tested across three difficulty levels, with a limited number of five attempts per test case. Performance is measured by the final success rate, averaged over five complete runs, reflecting both an understanding of Mermaid syntax and effective tool usage.