Pendulum tunes the personality of a local GGUF model with sliders. No endless prompt, no token to guess: it looks where each trait lives inside the model, and pushes there.

 

 

01 · MEASURE	02 · TUNE	03 · EMBODY	04 · EXPORT
A ten-axis profile, read inside the model: warmth, humor, sarcasm, vulgarity…	One slider per trait. A judge checks the model still makes sense.	A whole character: who it is, what it wants, who you are to it. Trained, not role-played.	One file for llama.cpp or LM Studio. The original model stays untouched.

 

Start

# PyTorch for your GPU: https://pytorch.org/get-started/locally/
pip install torch --index-url https://download.pytorch.org/whl/cu130
pip install -r requirements.txt
python Pendulum.pyw

On first launch, an assistant sets up the rest in four screens: working folder, llama.cpp (found, or downloaded for your machine), model folders, judge.

 

How it works

 GGUF model ─▶ lens ─▶ workspace band ─▶ personality directions ─▶ sliders
                                                                      │
              the guard scales back what derails ◀─ judge ◀─ answers ◀┘

A lens translates the inner layers into words. In it we find the band where the model thinks rather than writes. That is where each trait gets its direction, calibrated on real answers.

 

In numbers

83 %	Qwen3.5-4B becomes Bubulle, a goldfish, name included (judge score; on questions never seen in training: 75 %)
27 B	trained on two consumer GPUs, 16 + 12 GB: the weights stay quantized as in the GGUF
0.3 s	to apply the sliders as set, without the judge
6×	faster: the profile is read instead of generated (0.98 correlation with generation)

 

Requirements

• Windows 10 or 11, Python 3.11 or newer. Linux: should work, untested so far (or through Docker).
• An NVIDIA GPU, preferably.
• GGUF models. Safetensors are converted once.

 

Further

• Full guide · every window, every setting, every measure
• Code map · what lives where
• Scripted mode · python Pendulum.pyw --load model.gguf --lever warmth=-1,humor=0.8 --chat "Hello"
• Docker · docker compose up --build, for those who enjoy it

 

_{Jacobian lens: Anthropic · engine: llama.cpp}