arnomatic/Qwen2.5-3B-Instruct-heretic
arnomatic/Qwen2.5-3B-Instruct-heretic is a 3.1 billion parameter instruction-tuned causal language model, based on Qwen2.5-3B by Qwen, specifically modified using Heretic v1.1.0. This model is decensored, exhibiting significantly reduced refusal rates compared to its original counterpart. It retains the Qwen2.5 base model's enhanced capabilities in coding, mathematics, instruction following, and long-text generation up to 8K tokens, making it suitable for applications requiring less restrictive content policies.
Loading preview...
arnomatic/Qwen2.5-3B-Instruct-heretic Overview
This model is a decensored version of the Qwen/Qwen2.5-3B-Instruct, created using the Heretic v1.1.0 tool. It is designed to provide responses with significantly fewer refusals, as evidenced by a refusal rate of 3/100 compared to the original model's 97/100.
Key Capabilities Inherited from Qwen2.5-3B-Instruct
- Enhanced Knowledge & Reasoning: Improved capabilities in coding and mathematics due to specialized expert models.
- Instruction Following: Significant improvements in adhering to instructions and generating structured outputs, including JSON.
- Long-Context Support: Supports a context length of up to 32,768 tokens and can generate texts up to 8,192 tokens.
- Multilingual Support: Capable of processing and generating text in over 29 languages.
- Robustness: More resilient to diverse system prompts, enhancing role-play and chatbot condition-setting.
Abliteration Parameters
The decensoring process involved specific abliteration parameters, such as direction_index (22.56) and various attn.o_proj and mlp.down_proj weight adjustments, which contribute to its altered behavior.
When to Use This Model
This model is particularly suited for use cases where the original Qwen2.5-3B-Instruct's content moderation or refusal policies are too restrictive, and a more open, less censored response is desired, while still leveraging the strong base capabilities of the Qwen2.5 architecture.