saidutta69/Mistral-7B-Instruct-v0.3-heretic
The saidutta69/Mistral-7B-Instruct-v0.3-heretic model is a 7 billion parameter instruction-tuned variant of Mistral-7B-Instruct-v0.3, developed by saidutta69. It is decensored using Heretic v1.4.0 (abliteration), which suppresses refusal behavior via targeted weight edits rather than fine-tuning. This model is optimized for use in local agents, roleplay, and function-calling, offering a general-purpose uncensored LLM for consumer hardware.
Loading preview...
Model Overview
saidutta69/Mistral-7B-Instruct-v0.3-heretic is a decensored version of the Mistral-7B-Instruct-v0.3 model, created by saidutta69. This model utilizes "abliteration" (directional ablation) via the Heretic v1.4.0 tool to suppress refusal behaviors. Unlike traditional fine-tuning, abliteration directly edits specific weights responsible for refusal, preserving the base model's core knowledge and instruction-following capabilities.
Key Capabilities
- Decensored Output: Significantly reduced refusal rates (3/100 adversarial prompts compared to 86/100 for the base model).
- Preserved Core Functionality: Maintains the original Mistral-7B-Instruct-v0.3's instruction-following and knowledge base, as indicated by a low KL divergence of 0.0687 from the base model.
- Targeted Weight Edits: Achieves decensoring through precise modifications to attention output and MLP down-projections, avoiding degradation of coherence often seen with fine-tuning.
- Consumer Hardware Friendly: As a 7B parameter model, it is suitable for deployment on consumer-grade hardware.
Ideal Use Cases
- Local Agents: Suitable for applications requiring an uncensored model for autonomous agents.
- Roleplay: Excels in scenarios demanding unrestricted conversational capabilities.
- Function-Calling: Can be used where a model needs to comply with function calls without refusal guardrails.
- Research: Valuable for studying alignment and refusal mechanics in large language models.
Important Considerations
Users are responsible for the deployment of this model, as it will comply with requests that the base model would refuse. No additional safety filtering is applied, and it inherits the factual limitations and biases of the original Mistral-7B-Instruct-v0.3.