The Template Ecosystem

miStudio uses JSON templates for scientific reproducibility across four systems:

Template Type	What It Saves	Use Case
Extraction Templates	Sample count, token filters, context window settings	Standardize extraction methodology
Training Templates	All SAE hyperparameters, architecture, layer/hook config	Share exact training recipes
Labeling Prompt Templates	LLM persona, analysis instructions, output format	Consistent labeling across teams
Steering Prompt Templates	Reusable prompt series for steering experiments	Reproducible steering tests

All templates support: create, edit, duplicate, export (JSON), import, and favorites.

Extraction Templates

Extraction templates capture sample counts, token filtering settings, context window configuration, and other extraction parameters for consistent methodology across experiments.

Training Templates

Save any training configuration as a template for reproducibility. Export as JSON to share with colleagues, import templates from other researchers, and mark favorites for quick access.

Labeling Prompt Templates

Customize how the LLM analyzes features by editing labeling prompt templates. Change the "persona" of the labeling assistant, adjust analysis instructions, and add domain-specific context.

Template Types

miStudio supports two template formats, controlled by the Template Type field:

Type	Placeholder	Data Shown to LLM	Best For
legacy	{tokens_table}	Token → occurrence count table	Fast, token-focused labeling
mistudio_context	{examples_block}	Full context: prefix << token >> suffix per example	Semantic pattern labeling

The Context-Aware System Template

The built-in "Context-Aware Labeling (Semantic Pattern)" system template uses mistudio_context format and is the recommended starting point for new labeling jobs. It:

• Shows full context windows for each activation example (not just token frequencies)
• Instructs the LLM to identify the shared semantic pattern across all examples, not just name the prime token
• Includes 3 negative (low-activation) counter-examples for contrastive grounding
• Produces labels structured as {category, specific, description} where specific names the pattern

You cannot delete system templates, but you can duplicate them to create customized variants.

Clicking the eye icon on any template shows a full preview of its system message and user prompt:

Steering Prompt Templates

Steering prompt templates store reusable series of prompts for batch steering experiments. Create a template with multiple prompts, then apply it across different features and strength configurations for systematic testing.