Product · Download
PSE Analyzer
The anatomy of a model's weights. PSE Analyzer opens a GGUF file, measures the eigenvalue spectrum of every weight matrix, compares one precision of a model against another, and seals what it finds in a dossier. It runs on an ordinary CPU, and nothing leaves your computer.
Download
Windows needs nothing installed. Ubuntu needs Python 3.10 and one internet connection.
The Windows files carry everything but torch. The Ubuntu package carries the program and builds its own Python environment on first run, so numpy and scipy arrive as wheels tuned to your CPU.
1d47c403b3219f70ae36b0b000e8158a59bbfb2efe2cd9cdb43a68ba3806852737f64f3d9224de86ed3ae3951b32da36d737241d02bb0251504b6f1bfe52e5c694577c6c2100dc4616d6665ca7854490f4317f31a1e2fe5ac7114ea81cce36c8Before you run it
These builds are not code-signed. On Windows, SmartScreen shows "Windows protected your PC": choose More info, then Run anyway. Verify the file first and compare the result with SHA256SUMS.txt from the same release.
# Windows (PowerShell)
Get-FileHash .\PSE-Analyzer-0.3.1-windows-x64.exe -Algorithm SHA256
# Ubuntu
sha256sum pse-a_0.3.1_all.debWhat it does
Weight anatomy first, activation physiology when you want it
Most of the tool is pure numpy and scipy and reads the file one matrix at a time, never the whole model into memory. Four missions need torch and read the model while it runs; they are handed to a torch environment kept alongside, which the tool can build for you. PSE Analyzer has protection levels 1 and 2 built in.
- doctor measures the machine first: CPU, RAM, BLAS speed, and what the model in front of it allows. No surprises in the middle of a run.
- l1 sweeps every weight matrix: Marchenko-Pastur bulk, heavy-tail exponent, separated directions, per tensor type, since a Q4_K_M file is a mixture of types rather than one format.
- qdiff compares two precisions of the same model, ideally against f16, and shows what a quantisation kept, moved or lost.
- lens, sae, jspace, ppl read the model while it runs: logit lens by layer, a feature dictionary for one layer, a map of J-space, perplexity along a precision ladder. These need torch.
- Every result is sealed into a dossier keyed by the model, with the configuration and seeds that produced it. The web view over the dossiers listens on 127.0.0.1 only.
Requirements
What you need on the computer
- A GGUF model file on the computer. The tool reads it one matrix at a time and never loads the whole model into memory.
- Windows: nothing else. Ubuntu: Python 3.10 or newer; numpy and scipy are installed into a user environment on the first run.
- torch only for lens, sae, jspace and ppl. The tool builds a torch environment alongside on request, or adopts one you already have.
- Disk for your data: harvests run to tens of gigabytes. Point PSE_A_DATA at a drive with room.
Licence and status
Open source, designed by Andrej Ristić
PSE Analyzer is free and open source under GPL-3.0, with its source at github.com/zeravatechnologies/pse-analyzer. The methods are Andrej Ristić's; the specification and plan ship with the Ubuntu package and are described in the documentation.
- 0.3.1 identifies a model by its header plus samples of its weights, so two models of one geometry no longer share a dossier. Older dossiers are carried over the next time their model is opened.
- Data stays on your computer: %LOCALAPPDATA%\pse-a on Windows, ~/.local/share/pse-a on Ubuntu. The web view listens on 127.0.0.1 only.
Contact
Found something in a spectrum?
Tell us which model and which precision. Andrej reads these.
Research inquiries: see the published papers · NDA available for commercial discussions · Toronto, Ontario, Canada