llmfan46/Qwen3.5-27B-Writer-V2-uncensored-heretic
The llmfan46/Qwen3.5-27B-Writer-V2-uncensored-heretic is a 27 billion parameter Qwen3.5-based language model, fine-tuned for creative writing and translation. This version significantly reduces content refusals by 91% (8/100 vs 93/100) compared to its original base model, ConicCat/Qwen3.5-27B-Writer-V2, while maintaining quality with a low KL divergence of 0.0274. It aims to improve writing quality and fluidity, addressing the original model's stiffness in creative and translational outputs.
Loading preview...
Overview
llmfan46/Qwen3.5-27B-Writer-V2-uncensored-heretic is a 27 billion parameter model derived from ConicCat/Qwen3.5-27B-Writer-V2. This iteration focuses on enhancing creative writing and translation capabilities by addressing the original model's tendency for stiff outputs. A key differentiator is its 91% reduction in refusals (8/100 compared to the original's 93/100), achieved through a decensoring process using Heretic v1.2.0 with the Arbitrary-Rank Ablation (ARA) method, specifically targeting attn.o_proj and attn.out_proj components.
Key Capabilities
- Reduced Censorship: Achieves significantly fewer content refusals while preserving model quality, indicated by a low KL divergence of 0.0274.
- Improved Writing Quality: Fine-tuned to produce more natural and less rigid text for creative writing tasks.
- Enhanced Translation: Aims to deliver more fluid and less literal translations.
- General Intellect Retention: Despite specialized fine-tuning, it maintains general instruction following and intellectual capabilities through the inclusion of instruct data.
Training and Performance
The model was initially trained on a mix of instruct, roleplay, and writing data, followed by extensive training on book chunks. MMLU (Massive Multitask Language Understanding) test results show a slight decrease in overall accuracy (0.8469 vs 0.8562 for the original), indicating a minor trade-off for the significant reduction in refusals. Recommended settings include a ChatML template, temperature of 0.7, top_p of 0.95, and a moderate dry penalty of 0.4-0.8.