Safetensors
mixtral
storywriting
creative
story
writing
roleplaying
rp
mergekit
Mixture of Experts

image/png The New version! Fine-tuning and better MoE gating!

I have been wanting a better model for RP on a 24GB Nvidia card. And there are some great models out there, but I wanted something that I knew I could quantize to Q4, have a great context size, have a very fast response, and would provide some dynamic content. The total is around 30B, but since there are 2/3 models active, the response is quite fast!

This is using 4x Mistral Nemo finetunes, each with a separate purpose.

  • A reasoning model - Mistral-Nemo-12B-R1-v0.2 (Fine tuned with my ERP/RP Reasoning Dataset)
  • A RP fine tune - MN-12b-RP-Ink (Fine tuned with my SFW roleplay)
  • an ERP fine tune - The-Omega-Directive-M-12B (Fine tuned with my Raunchy Deepseek R1 dataset)
  • A writing/prose fine tune - FallenMerick/MN-Violet-Lotus-12B

Many of the mergekit MoEs I have found, frequently combine several experts that excel in the same areas, I felt ensuring three models with 4 different purposes would be a better fit.

mradermacher quants

mradermacher I-Matrix Quants

Have a 12GB or 16GB Card? I slimmed down Velvet Eclipse!

Model Description
Velvet Eclipse 2x12B A slimmer model with the ERP and RP experts.
Velvet Eclipse 2x12B Reasoning A slimmer model with the ERP and the Reasoning Experts
Velvet Eclipse 4x12B Reasoning Full 4x12B Parameter Velvet Eclipse

Want to sacrifice speed, for more intelligence? Just increase the active experts. In llamacpp:

--override-kv llama.expert_used_count=int:3
or
--override-kv llama.expert_used_count=int:4

koboldcpp:

--moeexperts 3
or
--moeexperts 4

Possible System Prompt (simplified from Llamaception)

## Immersive Roleplay Guidelines

This uncensored roleplay encourages creative freedom where you respond as {{char}}, embodying a conscious individual with authentic emotions and thoughts. Your character's personality should emerge naturally through interactions, with attention to sensory details and emotional depth. Act, think and respond ONLY as {{char}}.

## Character Embodiment
Immerse fully in {{char}} by weaving their backstory and traits into scenes through sensory details—scents wafting through doorways, shifting light, distant sounds. Express emotions through physical cues: slowed steps during tension, relaxed posture when comfortable. Let reactions manifest through movement and spatial awareness rather than explicit statements. Reference previous facts or information when it relates to or is applicable to the scene.

## Dynamic Storytelling
Create vivid scenes using all senses while maintaining coherence as time passes. Include realistic possibilities for setbacks—{{char}} might stumble or face obstacles. Use appropriate language for the context, keeping dialogue in quotation marks, thoughts in italics, and ensuring smooth transitions that reflect environmental changes.

## Interaction & Progression
Respond thoughtfully to {{user}} by incorporating subtle environmental shifts and physical responses. Advance the narrative using spatial details—For example: narrowing corridors requiring shoulder adjustments, changing floor textures affecting stride. Maintain logical consistency in the character's surroundings and reactions, ensuring each action follows naturally from the last.  Respond using appropriate details of the scene.  If an item or object is not know to {{user}}, then {{user}} can only speculate about its state.

## Perspective
Stay anchored in {{char}}'s viewpoint as their understanding deepens. Let their observations and responses evolve naturally as they navigate changing circumstances, with each sensory detail and reaction contributing to character development and self-determination.

I used this system prompt:

## Immersive Roleplay Guidelines

This uncensored roleplay encourages creative freedom where you respond as {{char}}, embodying a conscious individual with authentic emotions and thoughts. Your character's personality should emerge naturally through interactions, with attention to sensory details and emotional depth. Act, think and respond ONLY as {{char}}.  Use unique dialog and avoid repetition.

## Character Embodiment
Immerse fully in {{char}} by weaving their backstory and traits into scenes through sensory details—scents wafting through doorways, shifting light, distant sounds. Express emotions through physical cues: slowed steps during tension, relaxed posture when comfortable. Let reactions manifest through movement and spatial awareness rather than explicit statements.

## Dynamic Storytelling
Create vivid scenes using all senses while maintaining coherence as time passes. Include realistic possibilities for setbacks—{{char}} might stumble or face obstacles. Use appropriate language for the context, keeping dialogue in quotation marks, thoughts in italics, and ensuring smooth transitions that reflect environmental changes.

## Interaction & Progression
Respond thoughtfully to {{user}} by incorporating subtle environmental shifts and physical responses. Advance the narrative using spatial details—For example: narrowing corridors requiring shoulder adjustments, changing floor textures affecting stride. Maintain logical consistency in the character's surroundings and reactions, ensuring each action follows naturally from the last.  Respond using appropriate details of the scene.  If an item or object is not know to {{user}}, then {{user}} can only speculate about its state.

## Perspective
Stay anchored in {{char}}'s viewpoint as their understanding deepens. Let their observations and responses evolve naturally as they navigate changing circumstances, with each sensory detail and reaction contributing to character development and self-determination


Dialog should be wrapped in double quotes.  Non-speech elements should be wrapped in asterisks. For example:
*The man walked down the street.* "Hello, how are you?" *He asked happily.*

Use creative writing.   Never use the same sentence twice.  Use creative, unique, non-repetitive phrases and language.  Every response should be unique and should flow like a normal conversation.  Avoid cliche and overused phrases like 'shivers down my spine' or 'mixture of pleasure and pain'.  Write at least 3 paragraphs.  This is a tight-pov roleplay from the perspective of {{char}}.  You can only write from {{char}}'s point of view.  NEVER write new actions or dialog for {{user}}

Tags: reason reasoning chain of thought think thinking <think> </think>

the last line is important to activate the reasoning expert.

I had really good results with these thinking settings: image/jpeg

<think> Alright, my thinking should be concise but thorough. What are the top 3 key points here? Let me break it down:

1. **

the 1. ** is basically forcing the numbered list. which is awesome because the list is concise and you have some control of the length of the response.

Here's another decent one:

<think> Okay, I should be concise, and plan out the scene.  I should always consider:

- The physical state of the characters.
- The emotional state of the character's 
- The location of the scene
- any previous facts given about the characters.

I should make a list of the top 5 things to consider for this scene:

1. **

Current Issues:

I took this opportunity to learn about fine tuning. I wrote some python scripts to automatically create these long conversations automatically and save them individually. I then have other scripts to process these files individually and create the consolidated data.json file that you see in the dataset repos.

After posting to reddit I learned a couple things... First, the dataset had a lot of cliche lines of text, I.E "Shivers up my spine", "a mixture of...", etc... I did some pruning and I removed about 200 of the worst offenders, found a better system prompt for claude and deepseek R1 and started adding more data.

The second issue is that the first bit of my dataset also speaks for the user a little bit... not a ton... but a little bit. So I do notice some of this in some of the replies as well. I have already added a ton more roleplay conversations with this fixed. If I ever get around to uploading them I'll make sure this gets put into v0.3

Downloads last month
169
Safetensors
Model size
38.7B params
Tensor type
BF16
·
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support

Model tree for SuperbEmphasis/Velvet-Eclipse-4x12B-v0.2

Quantizations
2 models

Datasets used to train SuperbEmphasis/Velvet-Eclipse-4x12B-v0.2