Lunarin's API Configuration
Hello, my little fishes! This guide gathers the LLM and proxy settings I’ve refined for text‑based roleplay on Janitor AI. You’re not obliged to copy everything – experiment freely and find the balance that suits you best. Under each point I’ll leave links to the sources I used, so you can dive deeper and tune things to your own taste. 🌸
Introduction to Janitor LLM
Janitor LLM is the platform’s native language model. Its behaviour is shaped by a custom prompt, generation parameters, and advanced controls. The following sections walk you through each part.
Path to Settings
- Open any chat with a character.
- Tap the three horizontal lines in the top‑right corner.
- Go to API Settings → janitor.
- Jump between the tabs: Custom Prompt, Generation Settings, and Advanced.
Custom Prompt
This is my current prompt, taken from kolach3’s guide and extended with NPC creation and Markdown formatting rules.
The last two paragraphs (NPC creation and Markdown formatting) are my additions.
If you want a shorter version or extra perks, check the full guide:
kolach3 prompts
Generation Settings
| Parameter | Recommended value | Notes |
|---|---|---|
| Temperature | 0.7 – 1.0 |
I adjust this depending on how creative or grounded I want the reply to be. |
| Max tokens | 1000 |
Not strictly necessary, but prevents endless loops of repetitive phrases. |
Advanced Settings
| Parameter | Value | Rationale |
|---|---|---|
| Top K | 40 |
Average value from community tests. |
| Top P | 0.9 |
Average value from community tests. |
| Rep. Penalty | 1.1 |
Average value from community tests. |
| Freq. Penalty | 0 |
Known to work poorly – leave it at zero. |
Freq. Penalty
Most users report that the frequency penalty does more harm than good. It’s safest to keep it untouched.
Prefill and Forbidden words I left empty.
Command for Chat Memory
The command below forces the bot to generate a structured, token‑efficient summary for the Chat Memory. It follows the developers’ recommendations.
How it helps
The bot saves key facts in a compact format, preserving context without wasting tokens.
OOC Commands
Out‑of‑character commands let you control the narrative – skip time, correct the bot’s behaviour, or explicitly consent to intimate scenes (yes, Janitor sometimes needs a notarised permission :p).
Two collections cover almost every need, and you can use them as a base to craft your own:
Introduction to Proxy (Owl Alpha)
I’m a student, so I’m always on the lookout for free, high‑quality options. Owl Alpha recently appeared on OpenRouter and turned out to be a hidden gem – it has no censorship on violent or sexual themes (political topics might be restricted, but I haven’t tested that). The only catch: free OpenRouter accounts get 50 requests per day. You can work around it by switching between several accounts. 😉
Path to Settings
- Open any chat with a character.
- Tap the three horizontal lines in the top‑right corner.
- Go to API Settings → proxy.
- Move between Generation Settings and Advanced as needed. (Optionally, use the Global Prompt field – just make sure each proxy‑specific prompt is phrased in a way the model understands best.)
Importing Proxy Configuration
I’ve exported my Owl Alpha settings as a JSON file. To use it:
- Click Import in the proxy configuration panel.
- Paste the snippet below and add your own API key from OpenRouter.
The jailbreak prompt is almost identical to my Janitor LLM prompt – only the Markdown presentation differs slightly.
New to proxies?
If you need step‑by‑step help setting up another model (DeepSeek, ChatGPT, Gemini, etc.), consult the official guides:
Part One – Absolute Beginner’s Guide
Part Two – Absolute Beginner’s Guide
Generation & Advanced Settings
Here everything is nearly identical to Janitor LLM, with a few important tweaks:
| Parameter | Value | Why? |
|---|---|---|
| Temperature | 0.8 |
Slightly warmer for more creative prose. |
| Max tokens | 0 |
Other LLMs rarely suffer from infinite‑text bugs, so we remove the cap. |
| Context size | 36000 |
You can try 48000–64000, but excessive context may overwhelm the model and cause glitches. |
Context overload
Setting the context size too high can “fry” the model’s brain. Stick to 36000 unless you know exactly what you’re doing.
The same OOC commands and the Chat Memory prompt should work with proxy models, though I haven’t tested them exhaustively yet.
In my experience, Owl Alpha surpasses JLLM in several ways: it actively drives the plot, isn’t afraid of creative character actions, avoids permission‑asking loops (“may I look at you?”), and stays away from overused phrases like “came as close as possible”, “reached out but didn’t touch”, or “assessed {{user}} like a predator”. I still need to test how well it handles physical violence towards the persona, but so far it’s a goldmine.
Final Provisions
I hope this guide helps you configure the perfect husband, wife, or whatever character you dream of.
May it bring you joy, my little fishes! 🌊
