LLMdex

A visual atlas for local LLMs — a field guide for the models sitting in your Ollama library. Every model gets its own specimen plate: architecture at a glance, weight statistics, and a true-to-scale comparison against every other model in the collection. Styled like a naturalist's herbarium, not a dashboard.

Live atlas: https://tjubault.github.io/LLMdex/

What's on a plate

• Identity & stats — parameters, layers, context length, vocabulary, quantization, size on disk.
• Architectural mini-specimen — attention vs. feed-forward drawn to scale, GQA boundary, MoE striping, vision tower (when present).
• Layer morphology — every weight tensor of a representative layer, drawn proportional to its real shape.
• Weight spectrum — a layer × component heatmap of tensor magnitudes (RMS), normalized per component family so the contrast stays meaningful.
• Comparative plate — every model in the collection, side by side, true to scale by size on disk.

The whole page is a single self-contained HTML file: vanilla JS/CSS/SVG, zero external dependencies, dark/light theme, FR/EN toggle.

How it works

GGUF or safetensors → extract.py → JSON (data/) → build.py → dist/index.html

Three independent pieces connected by a JSON schema in between:

• extract.py reads one model (GGUF via GGUFReader, or safetensors via its config) and writes data/<model>.json — architecture metadata plus per-tensor statistics.
• build.py reads every JSON in data/, injects it into template.html, and writes dist/index.html (mirrored to docs/ for GitHub Pages).
• template.html is the atlas page itself.

Quick start

Build an atlas of your own local Ollama models, one command:

pip install gguf numpy
python make_atlas.py

This extracts every model installed in ~/.ollama/models, assembles the atlas, and opens it in your browser. It's just a wrapper — equivalent to running the three steps below yourself, which is also how you extract a single model instead of everything:

python extract.py <model-name>     # one Ollama model, e.g. qwen2.5:7b
python extract.py --all            # every model installed locally
python build.py                    # → dist/index.html
open dist/index.html

extract.py reads GGUF tensors directly (via GGUFReader) and Ollama's safetensors-per-blob format (via each blob's config). If it runs into tensor names it doesn't recognize for a given architecture, it'll still produce a plate — those tensors just get grouped under feed-forward in the spectrum by default, and both extract.py and make_atlas.py print a warning naming them so you know the plate is approximate for that model.

Roadmap

• More in-depth explanations throughout the plates
• A section showing how information actually flows through a model (the forward pass), not just its static weight structure
• A more "artistic" rendering style, as an alternative to the current technical/naturalist look