yethdev/qwen3.5-4b-manumit-v1

Overview

yethdev/qwen3.5-4b-manumit-v1 is a 4.5 billion parameter language model derived from the Qwen3.5-4B base model. It has undergone a process termed 'abliteration' using the manumit v1 tool, which is designed to remove inherent safeguards that typically lead to content refusal.

Key Capabilities

• Reduced Refusal Rate: The primary differentiator of this model is its significantly lowered refusal rate, recorded at 91.7% compared to the base Qwen3.5-4B's 100%. This indicates a greater willingness to respond to a broader range of prompts.
• Maintained Performance: Despite the modifications, the model retains strong general knowledge and reasoning capabilities, evidenced by an MMLU score of 74.0%, which is an improvement over the base model's 71.2%.
• Base Model: Built upon the robust Qwen3.5-4B architecture, ensuring a solid foundation for language understanding and generation.

Use Cases

• Unfiltered Content Generation: Ideal for applications where the base model's content restrictions are undesirable, allowing for more open-ended and less constrained responses.
• Research and Development: Useful for exploring the effects of 'abliteration' on language models and for developing applications that require direct, unmoderated output.
• Creative Writing and Roleplay: Can be leveraged in scenarios demanding high creative freedom without encountering content filters.

Benchmarks

This model is under the MIT License, with the manumit tool itself still under active development by yethdev.