LuxBannerVelvetCafe
Velvet_Cafe
MN-VelvetCafe-RP-12B-V2 Q4_K_M GGUF Version
Thanks for the feedback — it helped fix two main issues in the previous release
-
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:
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:
SamplerSettingsModelButton200
-
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
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:
image
Outdated bloated formats I skip:
- 1500–2000 token cards full of P-lists, rigid bullets, personality tables, etc.
- Example of dense P-list style:
image
- Modern models read standard plain text formatting just fine — no need for over-engineered lists anymore
How I actually start RPs in SillyTavern:
- Use /sys to send a narrative summary first (sets scenario before any messages)
- Let the character send the real first message
- If context limit pushes the /sys prompt out:
- Repurpose it into the Group Chat Scenario field
- Or condense into Author's Notes
ModelButtonIL200
- 📖 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:
Configuration
The following YAML configuration was used to produce this model:
1models:
2 - model: kyx0r/Neona-12B
3 - model: PocketDoc/Dans-PersonalityEngine-V1.3.0-12b
4merge_method: slerp
5base_model: kyx0r/Neona-12B
6parameters:
7 t:
8 - value: 0.2
9 - filter: self_attn
10 value: [0, 0.2, 0.4, 0.6, 0.8, 1]
11 - filter: mlp
12 value: [1, 0.8, 0.6, 0.4, 0.2, 0]
13dtype: bfloat16
14chat_template: "chatml"
15tokenizer:
16 source: "base"
