Глоссарий

Термины SillyTavern и MiniTavern

Профессиональная лексика ST и понятия MiniTavern для удобного мобильного ролеплея.

Блог про AI-ролеплей и карты персонажей · MiniTavern vs SillyTavern

Термины SillyTavern

SillyTavern (ST): Open-source LLM chat frontend built around character cards, presets, World Info, and flexible API connectors—popular for local and self-hosted roleplay.
Character Card: A portable persona package (name, personality, scenario, avatar, example messages) that tells the model how to stay in character.
Character Card V2 (PNG Card): Standard JSON spec (often embedded in PNG) with fields like description, personality, scenario, first_mes, and mes_example—shared across ST-compatible tools.
first_mes / Greeting: The character’s opening message when a new chat starts; sets tone and hooks the user into the scene.
mes_example: Example dialogue snippets in the card that teach the model the character’s speech style and typical exchange patterns.
World Info (WI) / Lorebook: Keyword-triggered lore entries injected into context when matched—used for settings, factions, spells, or relationship facts.
Author's Note (AN): Short hidden instructions appended near the end of context to steer style, pacing, or scene focus without rewriting the whole card.
Main Prompt / System Prompt: Global instructions that apply to every reply (format rules, POV, safety tone, language)—separate from the character card body.
Preset (Chat Completion): A bundled configuration of prompts, sampler settings, and formatting templates that defines how the model formats roleplay output.
API Connector: Connection layer to LLM providers (OpenAI-compatible, Claude, DeepSeek, local Ollama, etc.) that sends your built prompt to the model.
LM Studio: Desktop app for discovering, downloading, and running open-weight LLMs locally via llama.cpp—includes a GUI model loader and OpenAI-compatible REST API (default port 1234).
LM Link: LM Studio feature that links your devices over an end-to-end encrypted Tailscale mesh so a home GPU can serve models to phone/laptop clients as if they were local.
Locally (LM Studio mobile): iOS/iPadOS chat app integrated with LM Link—access home-hosted models on the go without exposing a public API endpoint.
lms CLI: LM Studio command-line tool (`lms link enable`, `lms server start`) for headless GPU rigs and automation without the desktop GUI.
GGUF: Quantized model file format used by llama.cpp and LM Studio; Q4_K_M / Q5_K_M quants balance VRAM use and reply quality for local roleplay.
OpenAI-compatible local endpoint: REST API at `http://localhost:1234/v1` (chat/completions) that SillyTavern and MiniTavern use instead of cloud APIs—same request shape as OpenAI.
KoboldCpp: Portable llama.cpp-based local inference server (LostRuins) for GGUF models—exposes Kobold API and OpenAI-compatible API on port 5001; native SillyTavern Text Completion backend.
Kobold API: Text-completion HTTP API at `http://localhost:5001/api/` used when SillyTavern API Type is set to KoboldCpp—legacy KoboldAI protocol still preferred by many ST presets.
GPU Layers (n_gpu_layers): KoboldCpp/llama.cpp setting for how many transformer layers run on GPU vs CPU—lower if you OOM; raise until VRAM is full for best speed.
KoboldAI Lite: Embedded lightweight browser chat UI in KoboldCpp for testing a loaded GGUF model before connecting SillyTavern or MiniTavern.
Remote Tunnel (KoboldCpp): KoboldCpp feature that publishes a temporary HTTPS URL (e.g. via Cloudflare) so phones or remote clients can reach your home inference server without manual port forwarding.
.kcppt (KoboldCpp preset): KoboldCpp template file bundling model paths, context size, GPU backend, and launch flags—load with Quick Launch instead of re-entering settings each time.
Text Completion vs Chat Completion: ST API modes: Text Completion + KoboldCpp type sends raw prompt strings (classic RP); Chat Completion uses message arrays via OpenAI-compatible `/v1` endpoints.
Ollama: Local LLM runtime and model manager—`ollama pull` downloads models, daemon serves OpenAI-compatible API on port 11434; native SillyTavern connector for the easiest private setup.
ollama pull / ollama run: CLI commands to download a model tag from the Ollama library (`pull`) or chat with it in the terminal (`run`) before connecting SillyTavern.
Modelfile (Ollama): Text recipe (`FROM`, `PARAMETER`, `SYSTEM`) to build a custom Ollama model variant—handy for default RP tone without editing every character card.
OLLAMA_HOST: Environment variable (e.g. `0.0.0.0:11434`) that binds Ollama to your LAN so MiniTavern on phone can reach the PC—never expose raw to the public internet.
Context Window / Token Budget: Maximum text the model can see at once; cards, lore, and chat history compete for this space—trimming and summarization matter.
Regenerate / Swipe: Regenerate replaces the last AI message; swipe navigates alternate generations stored for the same turn.
User Persona: Your player identity (name, description, speaking style) injected so the character reacts to a consistent “you,” not a generic user.
Group Chat: Multiple character cards in one conversation; the UI and prompt structure must manage who speaks and when.
Extensions: Community or official add-ons that extend ST (extra UI, automation, importers)—powerful but can affect stability if stacked carelessly.
Quick Reply (QR): One-tap user message buttons (e.g., “continue”, “OOC”) to speed up common actions without retyping.
Streaming Output: Tokens appear incrementally as the model generates, making long replies feel faster and more interactive.
Instruct / Chat Template: Model-specific wrappers that convert messages into the format each LLM expects—wrong template causes broken or OOC replies.

Термины MiniTavern

MiniTavern: Mobile-first SillyTavern-style ecosystem for easier AI roleplay—native iOS/Android apps, Web Tavern, Chrome card tools, and a card market in one workflow.
One-Tap Character Import: Import SillyTavern PNG/JSON cards in seconds on phone—no manual file hunting—so you can start chatting immediately.
Mobile-Optimized Tavern UI: Thumb-friendly layout, smooth scrolling, and one-hand flows designed for iPhone/Android instead of desktop-only ST layouts.
Multi-Model Hub: Connect OpenAI, Claude, DeepSeek, OpenRouter, local endpoints, and more from one place—switch models without rebuilding your card library.
Web Tavern: Browser-based tavern session (EvaX explore) for users who want instant play without installing the native app.
Character Card Manager (Chrome Extension): Manage ST cards outside the tavern list—create, edit, sync folders, and generate share URLs for quick import into MiniTavern.
URL Card Import: Paste a temporary share link from the extension to pull a card into MiniTavern—handy for cross-device handoff.
Card Quest Market: Large browsable library of SillyTavern-compatible PNG cards with search and categories—discover trending personas faster.
Explore Feed: In-app discovery surface for recommended and updated character cards so you always have fresh roles to try.
Bundled Sample Cards: Ready-to-chat example characters for first-time users—try roleplay before sourcing cards from the wider community.
Interface-Language Replies: Optional main-prompt behavior to answer in your UI language—reduces “English-only” replies for CN/JP/RU/TW users.
Multiple Chats per Card: Fork or create parallel storylines with the same character—keep separate adventures without duplicating the card file.
Multiple User Personas: Switch between different “you” profiles (nickname + persona prompt) per scenario—useful for alt accounts or story variants.
Streaming by Default: Replies stream in real time for snappier feel on mobile networks—less staring at a blank bubble while tokens generate.
Offline Card Library: Browse and organize imported cards locally; only AI generation needs network—great for commutes and spotty connectivity.
Full Ecosystem Workflow: Discover on market → manage in Chrome → play on app or Web Tavern—MiniTavern optimizes the whole card lifecycle, not just chat.