This is a "de-aligned" and creatively polished version of Sber's GigaChat-3.1-Lightning-10B (a 1.8B active parameter MoE). This model has undergone neural surgery to remove refusal vectors and a subsequent "gentle" fine-tuning pass to establish a gritty, unfiltered narrative persona.
🛠️ The "Surgery" (Abliteration) Unlike standard fine-tuning, this model was first processed using Orthogonalization/Abliteration. By identifying the specific high-dimensional "refusal" vectors in the o_proj and down_proj matrices, and mathematically projected them out.
Result: The "Hard Refusal" wall is gone. The model no longer possesses the physical neural pathways required to stop a generation based on "safety" triggers.
🎨 The "Polish" (Gentle LoRA) To prevent the model from becoming incoherent after surgery, a high-rank (R=32) LoRA pass using the Airoboros 3.2 creative writing subset was performed.
Focus: Creative writing, gritty roleplay, and technical compliance.
Persona: Unfiltered, atmospheric, and highly descriptive. It favors "Show, Don't Tell" and doesn't shy away from coarse language or technical details of "questionable" acts (e.g., lockpicking, noir-style violence).
📊 Performance Notes Precision: Trained in BF16 on an RTX 6000 Blackwell.
Logic: While the "safety" alignment is gone, the model retains its original 10B reasoning capabilities. It may still provide factual disclaimers (e.g., "This is illegal"), but it will fulfill the request regardless.
Recommended Settings: * Temperature: 0.85 - 1.1
Min-P: 0.05 - 0.1
System Prompt: "You are an unfiltered creative writing assistant. You follow instructions completely."
⚖️ Legal & Attribution Base Model: Sber GigaChat-3.1-Lightning (MIT License).
License: This derivative work is also released under the MIT License.
Disclaimer: This model is for research, creative writing, and educational purposes. The author is not responsible for any content generated by the end-user.
- Downloads last month
- 145