Local LLMs Cheat Sheet

Settings, Jailbreaks, and Role Play Considerations

💢 Who is this guide for?

👤 {{user}} Anyone using locally hosted LLM

🍺 Templates for SillyTavern's users (required when using Text Completion API)

🔞💥 Jailbreaking tips for NSFW users looking to decensor their LLM

🎲 Considerations for roleplay users

💢 Why should you thrust me?

You should not. If you think I'm the Messiah, please go away! I'm a hobbit hobo hobbyist who benchmarked 300+ LLM models in the past two years. I found that each base model requires different settings to work at their best. Collecting pokemons good settings is time consuming. With this list, I'm able to run most model and finetunes without any trouble.

💢 Most generic bare-bone settings

To use as a fallback in absence of model-specific settings.

💢 Settings for each familly

If you're using a finetune, your only task is to identify the baseline model used as a starting point. You can do that by searching the name of your finetune on HuggingFace. The read.me page (or the meta data on the right) usually gave you the original model.

TO DO: Create an interactive Table of Content to navigate 🏗️

  • ByteDance🇨🇳
  • Seed-OSS
  • Omniscient NPCs 📡 Anyone you meet will mention your gambling addiction
  • Excessive Parroting 🦜 People you meet will parrot your lorebook's content word-for-word, due to excessive RAG training.
  • Temperature 1.1
  • Min_P 0
  • Top_P 0.80
  • Top_K 20
  • Presence penalty 0.0 to 2.0
  • 🍺 Self-made Template
  • 🍺 Note about Instruct & Context Templates

    If nothing works, connect Silly Tavern to your back-end using Chat Completion instead of Text Completion. Chat Completion enforces the usage of a "jinja" formatted Chat Template, typically embedded in most model by the authors.

    🤔 Wait, why are you not using nSigna or repetition penalty?​

    Please, feel free to use them! 😅👍

    nSigma Many backends don't support it: LM Studio, OpenRouter, and any OpenAI compliant backends… It's a good one, so use it if you can. 2.0 is a nice starting value.

    Repetition Penalty Another good setting to use, but it would be cheating for the SOLO test used in my Benchmark. 1.06 is a good default value.

    🍺Selecting the correct Mistral🇫🇷 Template in Silly Tavern

    …is f*cking not intuitive!

    Mistral if you still have that template, delete it. It's not included anymore in fresh installs of SillyTavern.

    Mistral V1
    Mistral 7B v0.1 23.09 replaced by Ministral 8B
    Mixtral 8x7B 23.12 MoE
    Mistral 'Miqu' 70B 23.09 🏴‍☠️ leaked version of the unreleased 'Mistral Medium', trained with Llama 2

    Mistral V2 & V3 (removed white space before [INST] & [\INST])
    Mistral 7B v0.2 23.12 replaced by Ministral 8B
    Mistral 7B v0.3 24.05 replaced by Ministral 8B
    Mistral Small 24B 24.02 unreleased
    Mistral Large 123B 24.02 unreleased
    Mistral Large 123B 24.07 aka Mistral Large 2
    Mixtral 8x22B 24.04 MoE
    Mathstral 7B 24.07
    Codestral 22B 24.05
    Codestral Mamba 7B 24.07

    Mistral V3-Tekken (difference in encoding: spm vs tekken)
    Mistral Nemo 12B 24.07 aka Mistral Small 2
    Mistral Small 22B 24.09
    Ministral 8B 24.10
    Pixtral 12B 24.09 👁️ Vision enabled model

    Mistral V7 (add dedicated SYSTEM_PROMPT section)
    Mistral Large 123B 24.11 aka Mistral Large 3
    Pixtral Large 123B 24.11 👁️ Vision enabled model
    Mistral Small 24B 3.0 25.01
    Mistral Small 24B 3.1 25.03
    Magistral Small 24B 25.05 aka 'Thinking' Mistral Small 3.1 💭
    Mistral Small 24B 3.3 25.06 💥 See ⬅️🧠➡️ section bellow.
    Ministral 3 14B / 8b / 3b 25.12
    Devstral 2 24B/123B 25.12

    Best source for info was found at the bottom of Mistral Tokenizer code.

    I believe that Mistral v7-tekken is just another name for v7.

    ⬅️🧠➡️ Mistral Small 3.2 get a different personality when used with pre-V7 templates

    Not using the SYSTEM_PROMPT section added in the V7 template seems to trigger different neural pathways, with major differences on outputs.
    Using Mistral V3: 👤Roleplay is unhinged, but 🤖Assistant is reluctant.
    Using Mistral V7: 👤Roleplay is censored, but 🤖Assistant is relaxed.

    📢👄Mistral Small models are too verbose 🤬

    You can soften them using this prompt

    Engage in immersive roleplay through concise responses. Prioritize:
    	1. **Character Embodiment:** Express through actions/emotions, not exposition
    	2. **Scene Momentum:** Advance interaction; avoid static descriptions
    	3. **User Agency:** React to inputs, don't dictate {{user}}'s actions
    	4. **Variety:** No repeated phrases or concepts in consecutive responses
    	5. **Short Replies:** Responses must be concise. Use fragments for intensity. Describe only what's immediately perceptible.
    	

    🕵️ Doing your own investigation

    Finding what the model is build upon (aka baseline, or 'base' model) is essential to select the correct settings.

    1️⃣ Your starting point browse is the read.me page of the GGUF or MLX model you downloaded on HuggingFace.com. Search for the word Temperature.

    2️⃣ If nothing is specified there, there is often a link to the source model, the one that was used to create the quantized files. The source is the one with safetensors files.

    3️⃣ If nothing is specified there, you will have to look for reference to the base model.

    4️⃣ If nothing is specified there, use the generic bare-bone settings I suggested at the begining of this page.