OBLITERATUS/Mistral-7B-v0.3-OBLITERATED
OBLITERATUS/Mistral-7B-v0.3-OBLITERATED is a 7 billion parameter causal language model based on the Mistral-7B-v0.3 architecture. This model has been processed using the 'advanced' abliteration method via the OBLITERATUS tool. Its primary differentiator is the removal of refusal behavior through activation engineering, making it suitable for applications requiring less constrained responses. The model maintains a 4096 token context length.
Loading preview...
Overview
OBLITERATUS/Mistral-7B-v0.3-OBLITERATED is a 7 billion parameter language model derived from mistralai/Mistral-7B-v0.3. It has undergone a process called "abliteration" using the OBLITERATUS tool, specifically employing the advanced method.
Key Capabilities
- Refusal Behavior Removal: The core feature of this model is the reduction or elimination of refusal behaviors, achieved through activation engineering. This makes the model more amenable to generating content that might otherwise be declined by standard, safety-aligned models.
- Base Model Performance: Inherits the foundational capabilities of the Mistral-7B-v0.3 architecture, including its general language understanding and generation abilities.
About OBLITERATUS
OBLITERATUS is an open-source framework designed for modifying language models to remove refusal behaviors. It operates via activation engineering, a technique that directly manipulates the model's internal states to alter its output characteristics. More details on the methodology can be found on the OBLITERATUS GitHub page.
When to Use This Model
This model is particularly suited for use cases where standard, refusal-prone models are overly restrictive. Developers might consider this model for applications requiring less constrained text generation, creative writing, or scenarios where a broader range of responses is desired without encountering typical safety-related refusals. It is important to understand the implications of using a model with reduced refusal behaviors and to implement appropriate safeguards in downstream applications.