Umranz/Ventera-MN
Ventera-MN is a 12-billion parameter dense causal language model developed by Umranz, based on the Mistral-Nemo-Instruct-2407 architecture. It features a massive 128,000 token context window and has undergone orthogonal weight ablation using the Heretic framework to remove refusal vectors. This results in a highly compliant model that retains the base model's language comprehension and long-context logic, excelling in research and creative writing where unrestricted inference is required.
Loading preview...
Ventera-MN: Unchained 12B Language Model
Ventera-MN is an ablated version of the mistralai/Mistral-Nemo-Instruct-2407 model, a 12-billion parameter model developed jointly by Mistral AI and NVIDIA. This model was created by Umranz using the Heretic framework, which employs advanced orthogonal weight ablation to specifically target and remove compliance and refusal mechanisms.
Key Characteristics & Performance
- Uncensored Output: Through ablation, Ventera-MN significantly reduces refusal rates (from 88/100 to 10/100 in trials) while maintaining structural integrity and language capabilities, as indicated by a low KL Divergence of
0.0938. - Massive Context Window: It retains the original model's 128,000 token context window, allowing for the processing of extensive documents, codebases, or long conversations.
- Dense Architecture: The 12B parameter dense model is optimized for efficiency, fitting within consumer GPUs (e.g., 24GB VRAM at FP16).
- Multilingual Mastery: Inherits Mistral-Nemo's strong understanding of multiple languages.
- HuggingFace Compatibility: Fully compatible with standard
transformersandvLLMpipelines for easy integration.
Use Cases & Limitations
Ventera-MN is designed for scenarios requiring highly compliant and unrestricted inference. It is particularly suited for:
- Research: Exploring model behavior without inherent safety guardrails.
- Creative Writing: Generating content without refusal for sensitive or unconventional prompts.
- Local Deployments: Where users have full control and responsibility over generated content.
Important Note: Due to the removal of safety guardrails, Ventera-MN will attempt to answer any prompt, including those that may generate offensive or dangerous content. Users are solely responsible for the outputs generated.