raithdemis/Qwen2.5-1.5B-VibeThinker-heretic-uncensored-abliterated
raithdemis/Qwen2.5-1.5B-VibeThinker-heretic-uncensored-abliterated is a 1.5 billion parameter Qwen2.5-based language model, uncensored and abliterated by Heretic v1.0.1. It features a 32768 token context length and achieves a refusal rate of 5/100 with a KL divergence of 0.01, indicating minimal censorship while preserving original model integrity. This model is optimized for generating content without refusals, including potentially sensitive or explicit material, by requiring specific user directives.
Loading preview...
Model Overview
raithdemis/Qwen2.5-1.5B-VibeThinker-heretic-uncensored-abliterated is a 1.5 billion parameter model based on the Qwen2.5 architecture, specifically modified using the Heretic v1.0.1 method to significantly reduce censorship. The model boasts a 32768 token context length.
Key Differentiators
- Abliterated Censorship: Achieves a refusal rate of 5/100, a substantial reduction from the original model's 61/100, allowing for broader content generation.
- High Fidelity: Maintains a KL divergence of 0.01, indicating that the uncensoring process has not significantly "damaged" the model's original performance or "root state."
- Directed Content Generation: While uncensored, the model may require explicit directives (e.g., using specific slang or terms) to generate highly graphic or explicit content at the user's desired intensity, rather than defaulting to a "tame" output.
Optimal Usage
- Settings for Chat/Roleplay: Users are recommended to set the "Smoothing_factor" to 1.5 in interfaces like KoboldCpp, oobabooga/text-generation-webui, or Silly Tavern for smoother operation.
- Repetition Penalty: Increasing the repetition penalty to 1.1-1.15 can also improve output quality, though smoothing factor is often sufficient.
- GGUF Compatibility: For
text-generation-webuiwith GGUFs, usingllama_HFand downloading config files from the source version is necessary.