// technical write-up · llm behavior
lobot: a local workbench for editing LLM behavior
lobot finds single neurons and steering directions inside open language models. It pushes them at a chosen dose and measures the change with a local judge. Everything runs locally on a Mac, and the safety guard always stays on. This write-up covers what lobot edits, how it checks each edit, and where Muse Spark fits.
Solo builder · runs locally on a Mac
At a glance
- type
- Technical write-up
- models
- Llama 3.2 3B · Qwen
- edits
- single neurons · steering directions
- checks
- local judge · held-out report cards · safety guard
- agents
- Muse Spark
- runs on
- local Mac
- live demo
- lohuunar-lobot.static.hf.space
- repository
- github.com/meschinidino/lobot ↗
On this page ＋Close ×
At a glance ＋Close ×
- type
- Technical write-up
- models
- Llama 3.2 3B · Qwen
- edits
- single neurons · steering directions
- checks
- local judge · held-out report cards · safety guard
- agents
- Muse Spark
- runs on
- local Mac
- live demo
- lohuunar-lobot.static.hf.space
- repository
- github.com/meschinidino/lobot ↗
What lobot edits
Retraining is the usual way to change how a model behaves. lobot takes a different path. It edits behavior from inside the model, and it leaves the training process out.
The workbench finds single neurons and steering directions inside an open model. It then pushes them at a chosen dose. The models in scope are Llama 3.2 3B and Qwen.
How each edit gets checked
Local judge
A local judge measures the change after each push. The result is a measurement, not an impression.
Held-out report cards
Report cards on held-out cases show whether the edit holds on cases the search did not use.
Safety guard
A safety guard checks each edit. The guard always stays on.
Where Muse Spark fits
Plain-words explanations
Muse Spark explains every result in plain words.
Second-opinion judge card
A second judge card gives a second reading of the same edit.
Dose-tuning agent
One agent tunes the dose for a push.
Neuron-search agent
One agent searches for the neurons to push.
Runs locally
Everything runs locally on a Mac. No retraining step sits between an edit and its measurement.
That keeps the loop short: find, push, measure, check. The home page card links to the live demo, and the announcement post shows the demo in motion.