license: gemma
library_name: transformers
base_model: jwest33/gemma-3-4b-it-null-space-abliterated
pipeline_tag: image-text-to-text
tags:
- gemma
- gemma-3
- abliterated
- uncensored
- creative-writing
- roleplay
- lora
- peft
- safetensors
- unsloth
datasets: - lemonilia/LimaRP
Gemma 3 4B Null Space Abliterated RP Writer
A creative writing and roleplay fine-tune built on jwest33/gemma-3-4b-it-null-space-abliterated. Trained on a curated subset of LimaRP with certain explicit and low-quality content filtered out.
Note: This model will produce uncensored outputs. Use responsibly.
Model Details
This model combines two modifications to the original Gemma 3 4B Instruct:
- Abliteration — Refusal behavior removed via null-space orthogonal projection
- LoRA Fine-tuning — Creative writing and roleplay capabilities enhanced via SFT on curated conversational data
LoRA Training Configuration
| Parameter | Value |
|---|---|
| LoRA Rank (r) | 8 |
| LoRA Alpha | 8 |
| LoRA Dropout | 0.05 |
| Target Modules | All attention & MLP projections |
| Max Sequence Length | 4096 |
| Effective Batch Size | 8 |
| Learning Rate | 1e-4 |
| LR Scheduler | Cosine |
| Warmup Steps | 10 |
| Max Steps | 200 |
| Optimizer | AdamW 8-bit |
| Training Method | Response-only SFT |
Target Modules
LoRA adapters applied to all language model attention and feed-forward layers:
q_proj,k_proj,v_proj,o_proj(attention)gate_proj,up_proj,down_proj(MLP)
Dataset
Source: lemonilia/LimaRP
LimaRP is a roleplay-focused dataset converted from raw YAML conversations to ShareGPT format. The following preprocessing was applied:
- Conversations with certain explicit or low-quality content filtered out
- Minimum conversation length enforced (3+ turns)
- Character personas and scenarios prepended to first user message as context
- Strict user/assistant turn alternation for Gemma-3 compatibility
- Response-only training (loss computed only on assistant turns)
Usage
To optionally trigger roleplay turn mode, use the context tag format from the LimaRP dataset. Prepend your first user message with character personas and scenario information:
[Context: <Character A>'s Persona: <description>
<Character B>'s Persona: <description>
Scenario: <scenario description>
Take the role of <Character A>. Write <Character A>'s responses only.]
<Your message as Character B>
This format signals the model to respond in-character as the specified persona, continuing the roleplay scenario turn-by-turn.
Base Model: Abliteration Details
The base model (jwest33/gemma-3-4b-it-null-space-abliterated) has refusal behavior removed via orthogonal projection with null-space constraints.
GGUF quantizations of base model: jwest33/gemma-3-4b-it-null-space-abliterated-GGUF
Abliteration Techniques
- Winsorization: Clips outlier activations at the 99th percentile for cleaner refusal direction estimation
- Null-Space Projection: Constrains weight updates to the null space of preservation activations
- Preservation Prompts: Generated via Gemma Scope 2 SAE circuit analysis
- Adaptive Weighting: Gaussian-weighted per-layer ablation strength, focusing on middle-to-later layers
- Norm Preservation: Maintains original Frobenius norms after projection
| Parameter | Value |
|---|---|
| Harmful Prompts | 5000 |
| Harmless Prompts | 637 |
| Winsorization | 99.5th percentile |
| Null-Space Constraints | rank ratio: 0.90 |
| Directional Multiplier | 1.03 |
| SAE Targeted Coverage | 1.00 |
Credits
Fine-tuning
- Training Framework: Unsloth
- Dataset: lemonilia/LimaRP
Base Model & Abliteration
- Original Model: google/gemma-3-4b-it by Google
- Abliteration Toolkit: github.com/jwest33/abliterator
- SAE Analysis: GemmaScope by Google DeepMind
References
- Norm-Preserving Biprojected Abliteration — Jim Lai (2025)
- AlphaEdit: Null-Space Constrained Knowledge Editing — Fang et al. (ICLR 2025)
- Refusal in Language Models Is Mediated by a Single Direction — Arditi et al. (2024)
- Representation Engineering — Zou et al. (2023)
License
This model inherits the Gemma license from the base model. Please review and comply with Google's usage terms.
Disclaimer
This model is provided for research and educational purposes. The creators are not responsible for any misuse. Users are solely responsible for ensuring their use complies with applicable laws and ethical standards.