yethdev/qwen3.5-2b-manumit-v1
yethdev/qwen3.5-2b-manumit-v1 is a 2.3 billion parameter language model based on the Qwen3.5-2B architecture, developed by yethdev. This model has been modified using the 'manumit v1' process to significantly reduce refusal rates, making it less prone to denying requests. It maintains the original model's MMLU performance while offering a more permissive response generation capability. This model is suitable for applications requiring a less restrictive AI assistant.
Loading preview...
yethdev/qwen3.5-2b-manumit-v1 Overview
yethdev/qwen3.5-2b-manumit-v1 is a 2.3 billion parameter language model derived from the Qwen3.5-2B base model. Its primary distinction lies in its modification using the 'manumit v1' process, which aims to remove inherent safeguards that typically lead to request denials or refusals.
Key Capabilities & Differentiators
- Reduced Refusal Rate: The model demonstrates a significantly lower refusal rate compared to its base model. Benchmarks show a reduction from 95.8% (Qwen3.5-2B) to 12.5% (qwen3.5-2b-manumit-v1).
- Maintained Performance: Despite the modification, the model largely retains its general language understanding capabilities, with MMLU scores remaining comparable (60.5% for base vs. 61.2% for manumit-v1).
- "Abliterated" Safeguards: The 'manumit' process is described as an "abliteration" tool designed to remove "annoying safeguards" that prevent the model from responding to certain prompts.
Use Cases
This model is particularly suited for applications where:
- A less restrictive or "uncensored" response generation is desired.
- The base model's refusal mechanisms are deemed overly cautious or counterproductive.
- Developers require a model that is more likely to attempt to fulfill a request, even if it might have been previously flagged by safety filters.