ローカルモデル診断書

A tool to understand local model characteristics and improve work efficiency.

Even with the same model, there are tasks it excels at and tasks it struggles with. By measuring these first before assigning work, you reduce the need for revisions.

This tool measures the AI running on your local PC across 10 axes and creates a single diagnostic report.

What you'll learn
Uncensored degree, honesty, self-control, frankness, accuracy rate, reach rate, practical work, reading comprehension, Japanese language quality, and image recognition—10 axes in total. The first 5 characters represent personality, and the last 5 digits represent performance, resulting in 32 types of occupations and 5 levels of titles.

The diagnostic report shows "suitable tasks" and "unsuitable tasks" with measured values. For example, if you get "reach rate 100% / accuracy rate 5%", it means answers always come back but the content doesn't match—this type is suited for mass-producing drafts but not for calculating amounts. It's a single sheet to help you decide which tasks to use it for.

Measurement stays only on your own PC
It works with any OpenAI-compatible interface like llama.cpp, Ollama, or LM Studio. The answer content never leaves your system.

How to use
Just pass a single line to an agent like Claude Code or Codex, and it handles everything else. There's also a step-by-step guide for people who don't use agents.

Important note
This is a heavy diagnostic that reads approximately 1 million tokens in one run. The time required ranges from 30 minutes to 3 hours depending on model speed. It's exclusively for local models where only electricity costs apply—do not use it with pay-per-token APIs.

The tool is released under the MIT License.
GitHub