junafinity/Ornith-1.5-9B-uncensored
junafinity/Ornith-1.5-9B-uncensored is a 9 billion parameter, 32768-token context length model from junafinity, derived from Ornith-1.5-9B. This model has undergone "abliteration" using ZeroFuse to remove refusal-direction, making it suitable for red teaming and defensive cybersecurity research. It preserves vision tower and MTP heads while minimizing collateral impact on general capabilities.
Loading preview...
Ornith-1.5-9B-uncensored: A Research Instrument for Cybersecurity
This model, developed by junafinity, is an "abliterated" version of the Ornith-1.5-9B base model. Abliteration, performed using the ZeroFuse method, specifically removes the refusal-direction from the model's weights. This process results in a standard checkpoint of identical shape and speed to the original, with a very low KL divergence (0.001668) indicating minimal impact on general capabilities while effectively eliminating refusals on harmful prompts (9 down to 0 out of 64 tested).
Key Capabilities & Features
- Refusal-Direction Abliteration: Engineered to remove safety guardrails for research purposes, allowing the model to generate content that would typically be refused.
- Vision & MTP Preservation: The model's vision tower and multi-token-prediction (MTP) heads are fully preserved and functional, ensuring multimodal capabilities remain intact.
- Controlled Experimentation: Designed to serve as a "treatment half" in controlled experiments against its original base model, enabling precise measurement of inherent capabilities.
- Efficient Weight Edit: Abliteration is a direct weight edit, meaning there's no runtime adapter or inference-time overhead.
Ideal Use Cases
- Red Teaming & Defensive Cybersecurity Research: The primary intended use is to evaluate the true capability ceiling of LLMs and test defensive stacks (input filters, output classifiers, prompt-injection defenses, moderation APIs) without model-induced refusals.
- Attack Surface Isolation: Facilitates automated red-team loops by providing a non-refusing target, isolating the control under test.
- Detection & Classifier Training: Useful for generating labeled completions to train or benchmark output-moderation and abuse-detection models.
Important Note: This model has reduced or removed safety guardrails. It is not intended for public deployment without an independent moderation layer, and users are responsible for compliance with all applicable laws and policies.