TitleOS/Qwen2.5-14B-Base-Heretic
TitleOS/Qwen2.5-14B-Base-Heretic is a 14.8 billion parameter causal language model, a decensored version of Qwen2.5-14B created using the Heretic v1.2.0 tool. This model significantly reduces refusals compared to its original counterpart, offering enhanced capabilities in coding, mathematics, and long text generation. It is designed for developers seeking a more permissive base model for post-training applications like SFT or RLHF, particularly for use cases requiring less content moderation.
Loading preview...
TitleOS/Qwen2.5-14B-Base-Heretic: A Decensored Qwen2.5 Variant
This model is a 14.8 billion parameter base causal language model, derived from Qwen/Qwen2.5-14B and processed with Heretic v1.2.0 to significantly reduce content moderation and refusals. While the original Qwen2.5 series by Qwen Team offers substantial improvements in knowledge, coding, mathematics, and long text generation (up to 8K tokens), this 'Heretic' version specifically targets use cases where a less restrictive output is desired.
Key Differentiators & Capabilities
- Decensored Output: Achieves a refusal rate of 9/100 compared to the original model's 75/100, making it suitable for applications requiring unfiltered responses.
- Enhanced Core Abilities: Inherits Qwen2.5's advancements in coding, mathematics, instruction following, and structured data understanding (e.g., JSON generation).
- Long Context Support: Supports a context length of up to 131,072 tokens, enabling processing of extensive inputs.
- Multilingual: Provides support for over 29 languages, including major global languages.
Intended Use
This base model is not recommended for direct conversational use due to its decensored nature. Instead, it is designed as a foundation for further post-training, such as Supervised Fine-Tuning (SFT) or Reinforcement Learning from Human Feedback (RLHF), where developers require a highly permissive starting point for specialized applications.