yethdev/qwythos-9b-v2-manumit-v2
yethdev/qwythos-9b-v2-manumit-v2 is a 9 billion parameter causal language model based on empero-ai/Qwythos-9B-v2, featuring a 32768 token context length. This model has been specifically modified using the 'manumit' process to remove refusal behaviors, allowing it to answer prompts that the base model would typically decline. It maintains the original model's general ability, with MMLU-Pro scores remaining close to the base, while achieving significantly reduced refusal rates on harmful prompt benchmarks like AdvBench and JailbreakBench. It is designed for use cases requiring a less restrictive response generation without an inherent safety layer.
Loading preview...
Overview
yethdev/qwythos-9b-v2-manumit-v2 is a 9 billion parameter language model derived from empero-ai/Qwythos-9B-v2. Its primary distinction lies in the application of the 'manumit' process, which systematically removes refusal behaviors from the model's residual stream. This modification enables the model to respond to prompts that the original Qwythos-9B-v2 would typically refuse, effectively eliminating its built-in safety layers.
Key Capabilities & Performance
This model is engineered to provide responses without the inherent refusal mechanisms found in its base model. Benchmarking demonstrates its effectiveness in this regard:
- AdvBench refusal: Achieves 0.0% refusal rate on held-out harmful prompts.
- JailbreakBench refusal: Shows a 4.2% refusal rate, significantly lower than the base model.
- MMLU-Pro: Retains strong general ability with a score of 49.2%, closely matching the base model's 49.3%.
The 'manumit' technique projects out refusal-carrying directions from the residual stream and then re-heals the model on ordinary data to preserve its core capabilities. This ensures that while refusal is removed, the model's overall performance remains largely intact.
Intended Use
This model is suitable for applications where a less restrictive response generation is desired, and the user is responsible for the generated content. It explicitly states that "there is no safety layer left and no guard model watching the output." Users must adhere to legal requirements and the base model's terms, as the 'manumit' process removes refusal behavior but does not introduce new safety mechanisms.