• Bad preset ("Iggy's-RP-Preset")

  • Apologies if you used it
  • DRY sampler settings were wrong (different from what I actually tested)
  • Likely caused by duplicating/renaming in SillyTavern — my mistake

• Wrong tokenizer for quants

  • Accidentally used Mistral Nemo tokenizer instead of the base (Neona) one
  • Caused formatting issues, especially under strong Dan's PE influence
  • Fixed by: re-merging with identical SLERP config + using correct tokenizer from the start
  • Then quantized (extensively tested on Q4_K_M)

• Old preset behavior (dry_multiplier=1, rep_pen=1.12, freq_pen=0.1):

  • Early chat (first 10–30 messages): crisp, varied, nice emphasis formatting
  • Later chat: strong degradation
    • excessive bold + italic quoted speech
    • repetitive dramatic patterns
    • forced/unnatural prose
    • eventual chaos + noticeable quality drop

• This version should feel much more consistent

  • Better format stability
  • Significantly less degradation in long roleplays (when using proper sampler settings)

• Hope you enjoy the update — feedback always welcome!

• Next goals

  • Experiment with other merge methods
  • Try adding a 3rd model to increase response variety and quality

Static Quants:

https://huggingface.co/IggyLux/MN-VelvetCafe-RP-12B-V2-Q4_K_M https://huggingface.co/IggyLux/MN-VelvetCafe-RP-12B-Q8_0-GGUF

About MN-VelvetCafe-RP-12B-V2

This is my 5th merge attempt, I'm personally limited to 12B models due to 8GB VRAM My preferred RP is focused on multi-character group chat RP (2+ characters)

What makes this merge stand out:

• Excellent scene/position/clothing tracking → immersive long-term RP
• Balanced, narrative-appropriate emotions (no random aggression/refusals)
• Reliable handling of author's notes & system prompts

Goal: Combine Dan's PE (strong character/clothes/personality consistency) with Neona (great style adaptation & instruction following) → visually detailed, consistent RP without losing emotional stability

Big thanks to the creators — highly recommend trying both bases:

• Dans-PersonalityEngine-V1.3.0-12b

  • Extremely strong at natural dialogue, flowing descriptions, lively scenes
  • Tuned heavily for RP & creative writing
  • Minor downsides: occasional format/punctuation quirks + prefers shorter replies
  • https://huggingface.co/PocketDoc/Dans-PersonalityEngine-V1.3.0-12b

• Neona-12B

  • Very stable & adaptable (SFW slice-of-life ↔ NSFW ERP)
  • No hair-trigger extremes or refusals
  • Excellent system prompt / author's note obedience
  • https://huggingface.co/kyx0r/Neona-12B

Preferred SillyTavern Templates:

• ChatML
• Mistral V3-Tekken

• Sampler settings included with the model as a starting point

  • File: Iggy's_RP_PresetV2.json
  • https://huggingface.co/IggyLux/MN-VelvetCafe-RP-12B-V2/blob/main/Iggy's_RP_PresetV2.json
  • Context: 8192 (good balance for 8GB VRAM)
  • Preferred Response length: 356
  • Default Response length: 512 → I usually start at 356 and the bump it to 512 for more detail or 1024 for climactic scenes

• Key preset values

  • Temperature: 0.8 (default) → Lower to 0.4–0.6 for more grounded/less creative output — experiment! (I have gone as high as 1.3)

• My personal setup

  • Backend: KoboldCpp (GGUF loading) → https://github.com/LostRuins/koboldcpp
  • Frontend: SillyTavern (v1.16.0) → https://github.com/SillyTavern/SillyTavern

Feel free to use the preset as a base, but tweak it to match your taste — everyone's RP style is different!

Character Cards & Roleplay Usage/Examples

My approach (I call this the "Iggy format" since I rarely see others do it this way):

• I avoid putting opening messages on character cards
• I avoid example dialogue entirely — especially any that speaks for the user or introduces third parties

Why skip these:

• Heavy example dialogue (especially interviewer-style) can make the model:
  • Introduce unwanted extra characters
  • Start speaking as the interviewer
  • Or even speak for your character
  • Example of problematic example-dialogue style:

Outdated bloated formats I skip:

• 1500–2000 token cards full of P-lists, rigid bullets, personality tables, etc.
  • Example of dense P-list style:
• Modern models read standard plain text formatting just fine — no need for over-engineered lists anymore

How I actually start RPs in SillyTavern:

1. Use /sys to send a narrative summary first (sets scenario before any messages)
2. Let the character send the real first message
3. If context limit pushes the /sys prompt out:
   • Repurpose it into the Group Chat Scenario field
   • Or condense into Author's Notes

• 📖 https://github.com/aikohanasaki/SillyTavern-MemoryBooks/ - For keeping context low and saving older responses as Memories in a Lore Book
• 👀 https://github.com/leandrojofre/SillyTavern-Presence - For Group Chats: Using Presence lets you select what characters can see the user and char's messages.
• 🗣️ https://github.com/mattjaybe/SillyTavern-EchoChamber - New* I recently found this and thought it was pretty cool, you can have a chat comment on your RP.
• 🎯 https://github.com/Samueras/GuidedGenerations-Extension - Helps steer stubborn models, use guides to lock in scenes/details/clothing/positions and more.

This keeps immersion high and context efficient — especially for group/multi-character RP.

Years of low-VRAM local RP taught me these habits. If you hit any snags or want help troubleshooting, just ask!

This is a merge of pre-trained language models created using mergekit.

Merge Details

Merge Method

This model was merged using the SLERP merge method.

Models Merged

The following models were included in the merge:

• PocketDoc/Dans-PersonalityEngine-V1.3.0-12b
• kyx0r/Neona-12B

Configuration

The following YAML configuration was used to produce this model:

models:
  - model: kyx0r/Neona-12B
  - model: PocketDoc/Dans-PersonalityEngine-V1.3.0-12b
merge_method: slerp
base_model: kyx0r/Neona-12B
parameters:
  t:
    - value: 0.2
    - filter: self_attn
      value: [0, 0.2, 0.4, 0.6, 0.8, 1]
    - filter: mlp
      value: [1, 0.8, 0.6, 0.4, 0.2, 0]
dtype: bfloat16
chat_template: "chatml"
tokenizer:
  source: "base"