AI Blocks › Research a topic properly
Model feature finder
Give it a language model. Get a small network trained to show which ideas its inner activity stands for.
Model feature finder is a free AI skill for the job "Research a topic properly". Give the instructions to an AI assistant that supports the required files and tools, then review its output.
Free · No account · MIT licence · Made by Orchestra Research · Level: expert · Thinking: high
Made by Orchestra Research. See the original on GitHub; we copied it as it is, with its licence.
What it does
A small helper network, trained on your language model, that shows which ideas its inner activity stands for. Instead of a blur of numbers, you get named features such as a place, a tone or a risky topic. The skill uses SAELens, a Python library. It picks one layer of the model, collects activations from lots of text, and trains a sparse autoencoder to split them into features where only a few fire at once. You then look at which text lights each feature up, and test it by turning it up or down. It needs Python, a GPU for larger models, and some comfort with research code.
You give it
- A language model and the layer to study
- A text dataset to read activations from
- A GPU and a Python setup
- A feature or topic you want to find (optional)
You get
- Training code for the feature network
- Named features with example text
- Charts of which features fire
- Steering and removal tests
- Notes on quality checks
Use it when
- You want to know what concept a layer of your model has learned.
- You are studying whether a model holds a feature for something unsafe.
- You want to steer a model by adjusting one named feature.
- You need to load a ready-made feature set and look around.
How to use it
- Add it to your AI once: download it and add the folder to Claude, ChatGPT or Codex as a skill.
- Ask for the job in your own words. Your AI opens the skill when it fits.
- It works best when your AI can run Python code on a computer with a GPU.