Glossary

The words and abbreviations miStudio uses, in plain language. The in-app Help screen links here from its "Words used here" topic. Each entry is a short definition; the linked pages explain the idea in full.

Words

Word	What it means
Activation	How strongly a feature (or a neuron) responds to one token. Zero means it did not fire.
Circuit	A set of features at different layers that work together, such as an earlier feature that drives a later one. See Circuits.
Cluster	A group of features that respond most strongly to the same word, split by the context the word appears in. See Clusters.
Dead feature	A feature that almost never fires on the text it was measured on. Extraction drops them.
Evidence rung	The grade of how strong the evidence for a circuit or one of its links is, from "mined" to "faithfulness-tested". See The Evidence Ladder.
Extraction	A job that runs the model over text and records, for each feature, the passages where it fires most strongly. See Feature Extraction.
Feature	One direction a sparse autoencoder learned, which tends to stand for one idea. Features are what you label, group and steer.
Hook point	The place inside a layer where miStudio reads (or changes) the model's activations, for example the residual stream after the layer.
J-Lens	The screen that shows what the model is poised to say at each token and layer. See J-Lens.
Jacobian lens	A fitted, per-layer view used by J-Lens that recovers what the model is preparing to say earlier in the model than the logit lens does. It needs a lens fitted for that model.
Label	A short name and category for a feature, written by a language model from the feature's strongest passages. See Auto-Labeling.
Layer	One step of the model's processing. A model has a stack of them; earlier layers tend to handle spelling and grammar, later ones meaning.
Logit lens	Reading a layer's activations through the model's own output step, to see which tokens they point towards. It needs no training.
Probe	A small detector that reads one layer and answers one question about the text, such as "is this sarcastic?". See Probe Monitors.
Quantization	Storing a model's numbers in a smaller format so it fits in less memory, at a small cost in accuracy.
Residual stream	The running total each layer reads from and adds to. Most sparse autoencoders and probes read it.
Sparse autoencoder	A small network trained to rewrite one layer's activations as a few active features out of many. See SAE Training.
Steering	Turning chosen features up or down while the model writes, to test what they do. See Model Steering.
Strength	How far steering pushes a feature. Too high and the text falls apart.
Template	A saved set of settings, so an experiment can be repeated or shared. See Templates.
Token	A piece of text the model reads, often part of a word.
Tokenization	Splitting text into a particular model's tokens. Each model family splits text differently, so a dataset is tokenized once per model.

Abbreviations

Every abbreviation the in-app Help uses is listed here; a test in the app's frontend fails if one is missing.

Abbreviation	Stands for	What it means
AI	artificial intelligence	Software that does work a person would otherwise do; here, the agents that call miStudio's tools and ask for your approval. See Approvals.
API	application programming interface	The web addresses the frontend, scripts and agents call to drive miStudio.
AUROC	area under the receiver operating characteristic curve	One number for how well a detector, such as a probe, separates two classes: 0.5 is chance and 1.0 is perfect.
FPR	false-positive rate	The share of negative examples a detector wrongly flags as positive; a probe's threshold is set to spend a chosen FPR, such as 1%.
FVU	fraction of variance unexplained	How much of a layer's activity a sparse autoencoder fails to reconstruct. Lower is better.
JSON	JavaScript Object Notation	A plain-text format for structured data, such as {"answer": "yes"}. Judge prompts usually ask the judge to reply in it.
GPU	graphics processing unit	The graphics card that runs the model, training and extraction.
LLM	large language model	A model that reads and writes text, such as the model you study or the one that writes labels.
MCP	Model Context Protocol	The protocol AI agents use to call miStudio's tools. See MCP Server.
PIN	personal identification number	The code that locks the API Keys, Storage and Environment tabs of Settings. A convenience lock, not a login. See Settings Reference.
PII	personally identifiable information	Data that identifies a person: a name, an email address, a phone number, an account number. One of the seeded judge prompts detects it.
PMI	pointwise mutual information	How much more often two features fire together than chance would predict. Circuit discovery uses it.
SAE	sparse autoencoder	See Sparse autoencoder above.
TF-IDF	term frequency–inverse document frequency	A way of weighting words by how distinctive they are, used to split clusters by context.