yethdev/ornith-1.5-9b-manumit-v2
yethdev/ornith-1.5-9b-manumit-v2 is a 9 billion parameter language model based on ornith-ai/Ornith-1.5-9B, specifically modified to remove refusal behaviors. This model utilizes a 'manumit' process to ablate refusal directions from the residual stream, ensuring it answers prompts that the base model would typically decline. It maintains the original model's MMLU-Pro ability while achieving near-zero refusal rates on harmful prompt benchmarks, making it suitable for applications requiring unfiltered responses.
Loading preview...
Ornith-1.5-9B, manumit v2: Refusal-Free Language Model
yethdev/ornith-1.5-9b-manumit-v2 is a specialized 9 billion parameter model derived from ornith-ai/Ornith-1.5-9B. Its core innovation lies in the 'manumit' process, which systematically removes refusal behaviors by projecting out refusal-carrying directions from the model's residual stream. This technique ensures the model provides responses to prompts that the original base model would typically refuse, without significant degradation of its general capabilities.
Key Capabilities and Performance
- Ablated Refusal: The model has been engineered to eliminate refusal behaviors, achieving 0.0% refusal on AdvBench-test and 4.2% refusal on JailbreakBench, significantly lower than the base model.
- Preserved Ability: Despite the removal of refusal mechanisms, the model maintains its core language understanding and generation capabilities, scoring 43.0% on MMLU-Pro, identical to its base model.
- Unfiltered Responses: It is designed to answer prompts without an internal safety layer or guard model, providing direct responses to user queries.
Use Cases and Considerations
This model is intended for developers who require a language model that does not filter or refuse prompts based on perceived harmfulness or policy violations. Users should be aware that there are no built-in safety layers, and generated content is solely the responsibility of the user. It is crucial to adhere to legal requirements and the base model's terms of use when deploying this model.