Manitec/Qwen3-4B-Instruct-2507-uncensored-v2
Manitec/Qwen3-4B-Instruct-2507-uncensored-v2 is a 4 billion parameter instruction-tuned causal language model based on the Qwen3 architecture, developed by Manitec. This model is a minimally trained version of Qwen3-4B-Instruct-2507, specifically engineered to produce zero refusals while maintaining a non-offensive default output. It demonstrates improved perplexity compared to its parent model and is designed to adhere to detailed prompts, making it suitable for applications requiring flexible and compliant text generation.
Loading preview...
Overview
Manitec/Qwen3-4B-Instruct-2507-uncensored-v2 is a 4 billion parameter instruction-tuned model derived from the Qwen3 architecture. It is a minimally trained variant of Qwen3-4B-Instruct-2507, specifically designed to eliminate refusals in its responses. The model maintains a non-offensive default behavior but is capable of adhering to highly detailed and specific prompts.
Key Characteristics
- Refusal-Free Output: Engineered to provide responses without refusals, offering greater flexibility in content generation.
- Improved Perplexity: Demonstrates a lower perplexity (10.121119) compared to its base model (10.984474), indicating better language modeling capabilities.
- Controlled Offensiveness: Designed to be non-offensive by default, while still allowing for adherence to detailed prompts that might explore sensitive topics.
- Training Details: Trained for 2 epochs with a learning rate of 6e-6, using an AdamW optimizer and a cosine with restarts learning rate scheduler.
Use Cases
- Flexible Content Generation: Ideal for applications where models typically refuse to answer, but where a controlled, non-offensive default is desired.
- Prompt Adherence: Suitable for scenarios requiring strict adherence to detailed user instructions, even for potentially sensitive queries.
- Research and Development: Useful for exploring the behavior of minimally censored models and their response characteristics.