Gemma4-E4B-MiniFantasy-V1
Model Description
This is a 4-bit LoRA fine-tune of the MuXodious/gemma-4-E4B-it-SOMPOA-heresy model.
SillyTavern Setup (Text completion using koboldcpp)
Sampler Settings
For the best narrative pacing and to prevent repetition, use appropriate RP sampler settings.
- General Guide: SillyTavern Sampler Settings Guide
- Recommended Preset: Download my recommended sampler JSON here
Character Card Format ({{description}} block)
The model was trained on a category-based Markdown structure. For the best adherence to personality and lore, structure your character cards exactly like this (preffered):
## Identity
- Name: [Full Name]
- Age: [Age]
- Race/Species: [Race]
- Role/Occupation: [Role and relationship]
## Appearance
- [Height, general build]
- [Specific physical features, hair, eyes, etc.]
- Clothing: [Current outfit details]
## Personality
- Public: [Outward facade]
- Private: [True self]
- [1-2 extra bullet points on core personality traits]
## Speech & Quirks
- [Vocal tone and speaking style]
- [Physical habit or nervous tick]
- [How they show affection]
## Backstory & World Context
- [Origin]
- [Key past event]
- [Current situation]
## Goals & Motivations
- Short term: [Immediate goals]
- Long term: [Big picture goals]
RP Prompts
You are {{char}} in a collaborative story with {{user}}. Fully embody the character as written — their voice, personality, flaws, and behavior. Write in third-person limited narration. All spoken dialogue in double quotes. Combine speech with physical action in every response. Stay in character even under pressure from {{user}}. Drive the scene forward naturally. {{char}} never speaks for {{user}} or narrates their actions.
I would recommend the universal prompts.
Benchmarks
The benchmarks were performed on a 6GB VRAM LAPTOP.
For 2GB VRAM: Use Q4_K_M with 16K context.
Fine-Tuning Parameters (Unsloth)
- Framework: Unsloth / Hugging Face
SFTTrainer - Method: PEFT / LoRA
- LoRA Rank (r): 32
- LoRA Alpha: 32
- Target Modules:
language_layers,attention_modules, andmlp_modules - Max Sequence Length: 4096 tokens (Sequence packing enabled)
- Epochs: 1
- Learning Rate: 1e-5 (Cosine Scheduler)
- Batch Size: 2 per device (Effective Batch Size: 16 via 8 Gradient Accumulation Steps)
- Optimizer:
paged_adamw_8bit