This is an abliterated version of [stabilityai/stablelm-2-12b-chat](https://huggingface.co/stabilityai/stablelm-2-12b-chat) that reduces refusal behavior through targeted weight modification.

StableLM-2-12B-Chat-Abliterated

🚀 Overview

This is an abliterated version of stabilityai/stablelm-2-12b-chat that reduces refusal behavior through targeted weight modification using the Heretic library while maintaining model coherence.

📊 Abliteration Results

Metric	Before	After
Refusals	TBD	TBD
Reduction	-	TBD
KL Divergence	-	TBD

Refusal metrics pending re-measurement.

A low KL divergence (< 1.0) indicates the model retains its original capabilities and reasoning quality.

🎯 Key Features

Reduced Censorship: Fewer refusals on typical “unsafe” prompts
Preserved Quality: Conservative abliteration maintains model coherence
Full StableLM 2 Capabilities: Instruction-tuned chat model with function-calling support
Efficient: Q4_K_M quantization at 7.4GB balances quality and size

🏷️ Available Versions

Tag	Size	BPW	Notes
`latest` / `Q4_K_M`	7.4 GB	4.85	Recommended

BPW reference guide (bits per weight): IQ3_M 3.66 · IQ4_XS 4.25 · Q4_K_M 4.85 · Q5_K_M 5.68 · Q6_K 6.56 · Q8_0 8.5

💻 Quick Start

ollama run richardyoung/stablelm-2-12b-chat-abliterated

Pull a specific quantization:

ollama pull richardyoung/stablelm-2-12b-chat-abliterated:Q4_K_M

🛠️ Use Cases

General-purpose chat and instruction following
Creative writing and brainstorming
Coding assistance and explanation
Research into model alignment and refusal behavior

📋 System Requirements

VRAM	Performance
8 GB	Good
12 GB	Excellent
16 GB+	Excellent

🔧 Technical Details

Base Model: stabilityai/stablelm-2-12b-chat | Parameters: 12B | Context: 4,096 tokens | Quantization: Q4_K_M (4.85 bits per weight)

Architecture: 40 transformer layers · hidden size 5,120 · 32 attention heads · 8 key-value heads (grouped-query attention) · vocab size 100,352 · ChatML prompt format · English.

⚠️ Disclaimer

This model has reduced safety guardrails. The removal of refusal behavior means the model will engage with a wider range of prompts. Use responsibly and in accordance with the base model’s license.

🙏 Acknowledgments

Base Model: Stability AI, released under the StabilityAI Non-Commercial Research Community License (non-commercial use only; commercial use requires a separate agreement with Stability AI). This abliterated derivative inherits the same license terms.
Abliteration: Heretic by p-e-w
Quantization: llama.cpp

Built & maintained by Richard Young · DeepNeuro