Research companion

AI Model Inspector

The research companion to Adventures in AI. A free local instrument for the work in Books One, Two, and Three. Version 1.0.1 is available for macOS.

1 · Books

How the three books are organized

Books One, Two, and Three remain one research series. A later book is another collection, not a new instrument.

Books overview with Book One, Book Two, and Book Three. — Book One and Book Two state that no separate measurement record is included. Book Three names the Freshman subjects: BASE, F5-1, F5-2, F5-3, and F6.

2 · Book Three

Recorded measurements and directions

The published Freshman record stays readable when the checkpoint files are not on the computer.

Recorded Book Three results, with effective-update magnitudes and Chapter 5 direction comparisons. — Each adapter shows rank, alpha, scale, effective-update magnitude, and the magnitude relative to the matching BASE projections. The Chapter 5 rows compare named Freshman pairs. These values were not calculated from a file on this computer.

3 · Hugging Face

Discover, download, and local files

Search reads a public model list. A download starts only after files are selected. Import Local is for a file already on the computer.

Hugging Face discovery results for SmolLM2, with Discover, Downloaded, and Import Local visible. — The search does not download weights. Public models stay available without a token. A lookup is a request to Hugging Face, so it leaves the computer.

4 · Checkpoint Inspector

Tensors, stored weights, and effective update

A Safetensors file on the computer can be inspected directly. Stored weights and the effective update are separate measurements.

Checkpoint Inspector for Freshman F5-1, showing 192 tensors and separate stored-weight results for LoRA A and LoRA B. — The file reports its tensor count, layers, and modules. Stored weights keep LoRA A and LoRA B apart. Effective update is the separate measurement of the scaled change, and it is not the stored-weight table.

5 · Cloud Models

OpenAI, Anthropic, and Gemini metadata

Each provider has its own key. A listed model is metadata from that account. It is not an inspection of the model’s weights.

Cloud Models with OpenAI, Anthropic, and Google Gemini listed separately. Saved keys are not shown. — Saved keys remain in the macOS Keychain and are not shown. View Models reads what that account’s API exposes. The screen does not send an inference request.

What the instrument does

AI Model Inspector brings the research of Books One, Two, and Three together. It can inspect a Safetensors file you supply, look up an accessible Hugging Face model, and read public model metadata from OpenAI, Anthropic, and Google Gemini.

The software is free to download under the Apache-2.0 license. That license does not include third-party model weights. Version 1.0.1 is the public source release for macOS.

Back to Our Work Download AI Model Inspector (Free)

On GitHub, open Assets and download ai-model-inspector-*-macos-source.zip. Requires macOS, Python, and Node.js. Follow the installation instructions in the included README.

Free · Open Source · Version 1.0.1