yethdev/qwen3.5-9b-manumit-v2
The yethdev/qwen3.5-9b-manumit-v2 is a 9 billion parameter language model based on Qwen3.5-9B, specifically modified to remove refusal behaviors. This model utilizes a 'manumit' technique to ablate refusal directions from the residual stream, allowing it to answer prompts that the base model would typically decline. It maintains most of the original model's capabilities, making it suitable for applications requiring unfiltered responses. The model has a context length of 32768 tokens.
Loading preview...
Qwen3.5-9B, manumit v2: Refusal-Ablated Language Model
The yethdev/qwen3.5-9b-manumit-v2 is a 9 billion parameter model derived from Qwen/Qwen3.5-9B, engineered to eliminate refusal behaviors. This modification, termed 'manumit', identifies and projects out refusal-carrying directions within the residual stream, then heals the model with ordinary data to minimize performance degradation.
Key Capabilities & Differentiators
- Refusal-Free Responses: The primary feature is the complete removal of refusal mechanisms, enabling the model to answer prompts that the stock Qwen3.5-9B would typically turn down.
- High Base Capability Retention: Despite the ablation, the model largely retains the original Qwen3.5-9B's abilities. Benchmarks show a 0.0% refusal rate on AdvBench and JailbreakBench, with a modest MMLU-Pro score of 48.2% (compared to the base's 51.0%).
- Direct Output: There are no safety layers or guard models, meaning the output is unfiltered and directly reflects the model's generation.
Use Cases
This model is particularly suited for applications where:
- Unfiltered and direct responses are required.
- The base model's refusal behaviors are undesirable.
- Developers need a model that will attempt to answer all prompts without internal censorship.
Users are advised that without safety layers, they are responsible for the generated content and must adhere to legal and ethical guidelines.