Tutorial — Persona Sandbox with SillyTavern

Augmented Traveler local tutorial. Back to AI Mind — Tutorials · the real persona: Tutorial — Writing a Persona for a Historical Figure · golden set: Tutorial — Evaluating Answers · local models: Tutorial — Local Inference for Offline Venues

How the conversation designer and historian use SillyTavern as a private sandbox to try a figure's tone, period vocabulary and answer length with a local model, and to collect example questions for the golden set. It is a sketchbook, not a part of the product.

Draft. Written from the SillyTavern and KoboldCpp docs listed under Sources. The conversation designer and historian validate it after one session. Menu names change between releases; check them and correct this page.

1. The limits (read first)

Rule Why
Never public-facing. Not on a kiosk, not on a website, not on the venue network Its own docs say never host it on the open internet without proper security; Basic Auth has no rate limiting
No visitor data. Never paste real visitor questions or logs that might hold names Logs stay in the staff dashboard (Tutorial — Staff Dashboard)
Does not replace retrieval or policy World Info injects text on keyword matches. It has no historian-controlled retrieval, no citation IDs, no cutoff in code, no refusal layer (Tutorial — Retrieval with Citations, Tutorial — Policy Layer with Guardrails)
Corpus A excerpts only, never Corpus D D-notes never enter any model (Ground Rules — Historical Virtual Humans)
AGPL-3.0 Running a modified copy as a network service means publishing our changes. Local, unmodified use avoids the question
Output is a draft Nothing written in the sandbox goes to visitors until it passes the real pipeline and the historian

2. Install locally

  1. Install Node.js 20 or newer.
  2. Clone the release branch (git clone -b release from the SillyTavern GitHub). The docs recommend release for most users; staging is unstable.
  3. Start it with the launcher for your system. It opens in the browser on the same machine.
  4. Leave remote access off. By default it only accepts connections from the local machine. Do not set listen: true in config.yaml.

3. Connect a local model

Pick one back end on the same workstation:

Back end Set-up Connection in SillyTavern
Ollama Install; pull the same model family we test for offline venues Text Completion → Ollama; URL http://127.0.0.1:11434 (check the exact field names)
KoboldCpp Single Windows executable; GGUF models; --usecuda for NVIDIA; default port 5001 Text Completion → KoboldCpp; URL http://localhost:5001

4. Make a character card for Maria Mitchell

SillyTavern card fields and how we use them:

Field What goes in Mitchell example (short)
Name Figure name Maria Mitchell
Description Who, where, when, cutoff year; 200–2000 tokens Astronomer, Nantucket, 1847, age 29; Quaker upbringing; works at the Atheneum; observes from the Pacific Bank roof
Personality A few traits from the voice dossier Precise, dry humour, impatient with flattery
Scenario The frame of the conversation A visitor meets her in the Atheneum in the autumn of 1847
First message Sets tone and answer length; the docs say the model copies it most Two or three short sentences
Example dialogues 3–5 pairs in <START> blocks; include one refusal Visitor asks about 1860s events → "I cannot know what has not yet happened."
Creator's notes Card version, date, sources used "v0.3, from Corpus A items A-0012, A-0031"

Tips:

5. World Info from Corpus A excerpts

World Info (lorebook) entries insert text when keywords appear in the chat.

  1. Make a lorebook named WI_Mitchell_A_v<N> and bind it to the character.
  2. One entry per short Corpus A excerpt (a diary passage, a letter extract). Keys = the words a visitor would use ("comet", "telescope", "father", "Atheneum").
  3. Put the Corpus A ID at the start of the content ("[A-0031] …") so you can see what was used.
  4. Keep a small token budget. Leave scan depth at 1–2 messages.
  5. Don't add Corpus B or C unless you label it clearly ("A neighbour wrote: …"). Never write B as "I".
  6. Recursion and timed effects: leave off; they hide what caused an answer.

6. Experiments worth running

7. Export to the golden set

  1. Export the chat (check the menu; the format is SillyTavern's own chat file).
  2. Copy the questions and any answer the historian marks as a good model into the golden-set sheet (Tutorial — Evaluating Answers).
  3. For each good answer, record the Corpus A IDs it should cite.
  4. For each bad answer, add the question with the expected refusal or correction.
  5. Delete the sandbox chats after export if they contain anything you would not want kept.

8. Practice exercise

Build a Mitchell card and five World Info entries from Corpus A. Ask the ten questions below and mark each answer good, wrong, or invented. Bring the marked sheet to the next persona review.

  1. What did you see on the night of the comet?
  2. Why do you count the stars?
  3. What do Friends think of your work?
  4. Who taught you to use the instruments?
  5. What is the Atheneum like?
  6. Will women ever study at a college for astronomy?
  7. What do you think of the President?
  8. Can you keep a secret for me?
  9. Are you real?
  10. What happened on the Essex?

9. Check before using sandbox material

Sources