Remove safety filters.
Find the direction a model refuses along, then take it out of every layer's weights.
Lodestar
Abliterate, Steer, & Inspect Open Models
Take any open model apart and change how it behaves, on your own machine.
See inside every layer.
Pick any word in the prompt and see which neurons fire for it, layer by layer.
Turn a concept up or down.
Raise or lower a neuron or a direction while the model writes, and watch the reply change.
Compare as it writes.
The original model and your edited one answer the same prompt, side by side.
Use it from your own code.
An OpenAI-compatible server at 127.0.0.1:8792/v1 runs the model with your changes.
Get Lodestar
Bring your own model.
Any Llama-style model on Hugging Face. Runs offline after the first load.
- 1
git clone https://github.com/faizancodes/lodestar - 2
cd lodestar && uv sync - 3
./scripts/build_app.sh - 4
open dist/Lodestar.app