CCSSNE/llmfan46-Qwen3.5-27B-Writer-V2-uncensored-heretic
CCSSNE/llmfan46-Qwen3.5-27B-Writer-V2-uncensored-heretic is a 27 billion parameter language model based on the Qwen3.5 architecture, developed by llmfan46. This model is a decensored version of ConicCat/Qwen3.5-27B-Writer-V2, specifically fine-tuned to improve creative writing and translation quality by addressing stiffness in the original model's output. It significantly reduces refusal rates (8/100 vs 93/100) while maintaining a low KL divergence of 0.0274, indicating strong preservation of the original model's quality. Its primary strength lies in generating more fluid and less restrictive creative text and translations.
Loading preview...
Model Overview
This model, CCSSNE/llmfan46-Qwen3.5-27B-Writer-V2-uncensored-heretic, is a 27 billion parameter language model derived from the Qwen3.5 architecture. It is a decensored version of ConicCat/Qwen3.5-27B-Writer-V2, created using the Heretic v1.2.0 tool with the Arbitrary-Rank Ablation (ARA) method. The primary goal of this iteration is to enhance the creative writing and translation quality of the base Qwen3.5-27B model, specifically by making its output less "stiff" and more natural.
Key Differentiators
- Significantly Reduced Refusals: Achieves an 8/100 refusal rate compared to the original model's 93/100, indicating a substantial reduction in content restrictions.
- Preserved Quality: Maintains a low KL divergence of 0.0274, demonstrating that the decensoring process largely preserves the original model's performance and capabilities.
- Improved Writing & Translation: Fine-tuned to produce more fluid and natural-sounding creative writing and translations, addressing a noted stiffness in the base model.
Training & Performance
The original base model was fine-tuned on a curriculum learning setup, starting with lower quality roleplay data, then progressing to higher quality writing data. It was trained on a mixture of instruct, roleplay, and writing data for three epochs, followed by eleven epochs on a smaller dataset of book chunks. MMLU scores for this Heretic version show a slight decrease (0.8469) compared to the original (0.8562), indicating a minor trade-off in general knowledge for the uncensored and improved writing capabilities.
Recommended Use Cases
- Creative Writing: Ideal for generating stories, scripts, or other creative content where natural language flow and reduced restrictions are desired.
- Translation: Suitable for translating text where a more natural and less rigid output is preferred.
- Roleplay: Benefits from the reduced refusal rates, allowing for more open-ended and less constrained roleplaying scenarios.