Skip to main content

Frequently Asked Questions

Which models can I use?​

Any HuggingFace transformer. miStudio has no architecture whitelist — layer discovery introspects each model dynamically, so standard architectures (GPT-2, Llama, Gemma, Qwen, Phi…) and unusual ones (hybrid models like LFM2) all work. Models that ship custom code need the trust_remote_code checkbox at download time.

How much VRAM do I need?​

It depends on model size and SAE width, not on miStudio itself. Practical reference points:

  • Getting started (TinyStories + a ~1B model, 8× SAE): comfortable on 8 GB
  • Serious work (2–7B models, 8–32× SAEs): 24 GB (e.g., RTX 3090/4090) is the sweet spot
  • Bigger models: quantize to Q8/Q4 for extraction, and train SAEs from cached activations so the base model doesn't occupy VRAM during training

See the OOM sizing rules for the arithmetic.

Do I need an OpenAI account?​

No. Labeling works with any OpenAI-compatible endpoint — vLLM, Ollama, LM Studio, or another local server. OpenAI's API is supported as an alternative; pick per-pipeline in Settings. Everything else in miStudio (training, extraction, steering) runs fully locally.

Can I use SAEs I trained elsewhere? Can I take miStudio SAEs elsewhere?​

Both directions work through the SAELens community standard format (cfg.json + sae_weights.safetensors):

  • Inbound: download pre-trained SAEs (Gemma Scope and others) from HuggingFace, or import from disk — see SAE Management
  • Outbound: every Neuronpedia export includes a saelens/ folder that loads directly in SAELens/TransformerLens; trained SAEs can also be uploaded to HuggingFace

Where does my data live?​

Everything is local: model weights, datasets, activations, and SAE weights on the filesystem (/data/ by default); metadata, features, and labels in PostgreSQL. Nothing leaves your machine unless you explicitly push to Neuronpedia, upload to HuggingFace, or use the OpenAI labeling backend.

What's the difference between "activation extraction" and "feature extraction"?​

Two different pipeline stages that unfortunately share a word:

  1. Activation extraction (Models panel) runs the base model over a dataset and caches its raw internal activations
  2. Feature extraction (SAEs panel) runs a trained SAE over activations to find each feature's top activating examples

See the extraction pipeline concept page for the full picture.

Why are my job progress bars not moving?​

Progress streams over WebSocket with polling fallback — see Real-time Updates & Progress. Nine times out of ten a page refresh resolves it; behind a reverse proxy, check WebSocket upgrade forwarding.

How do I reset everything?​

Stop the stack, drop the PostgreSQL database, and clear the data directory (/data/). On next startup the backend recreates the schema via migrations. To reset only the Settings PIN, use the MISTUDIO_BYPASS_PIN=true recovery flow described in the Settings Reference.

Is there an API I can script against?​

Yes — the full REST API (~100 endpoints) plus WebSocket channels are documented in the Reference section. The interactive Swagger UI is also available at /docs on the backend port.