llmfan46/Q3.5-BlueStar-v2-27B-ultra-uncensored-heretic-v2

Hugging Face
VISIONConcurrency Cost:2Model Size:27BQuant:FP8Ctx Length:32kTool Calling:SupportedPublished:Mar 23, 2026License:mitArchitecture:Transformer0.0K Open Weights Warm

llmfan46/Q3.5-BlueStar-v2-27B-ultra-uncensored-heretic-v2 is a 27 billion parameter language model based on the Qwen3.5 architecture, developed by llmfan46. This model is a decensored version of zerofata/Q3.5-BlueStar-v2-27B, created using the Heretic v1.2.0 tool with Arbitrary-Rank Ablation (ARA) method, significantly reducing refusals (5/100) while preserving core model quality (0.0671 KL divergence). It is specifically optimized for roleplay (RP) and creative writing tasks, supporting both 'thinking' and 'non-thinking' modes.

Loading preview...

Model Overview

llmfan46/Q3.5-BlueStar-v2-27B-ultra-uncensored-heretic-v2 is a 27 billion parameter model derived from the Qwen3.5 architecture, developed by llmfan46. This version is a decensored iteration of zerofata/Q3.5-BlueStar-v2-27B, achieved through the application of Heretic v1.2.0 and the Arbitrary-Rank Ablation (ARA) method.

Key Differentiators & Capabilities

  • Significantly Reduced Refusals: Achieves a refusal rate of 5/100, a substantial reduction from the original model's 99/100, indicating a highly uncensored output. This is achieved while maintaining a low KL divergence of 0.0671, signifying strong preservation of the original model's capabilities.
  • Optimized for Roleplay & Writing: Specifically designed for roleplay (RP) and creative writing tasks, aiming to improve intelligence and creativity while addressing repetition issues.
  • Flexible 'Thinking' Modes: Supports both 'thinking' and 'non-thinking' modes, with the 'thinking' mode requiring a \n prefill as per its training.
  • Repetition Mitigation: Incorporates custom loss masking during training to reduce repetitive phrases and overused words, a common challenge in RP datasets.

Performance & Training

  • Capability Preservation: PIQA (Physical Intuition Question Answering) benchmark scores show close alignment with the original model, with acc_norm values indicating good preservation of reasoning abilities despite decensoring.
  • Training Details: Fine-tuned using Axolotl with approximately 27 million tokens, employing techniques like custom loss masking on RP datasets to enhance output quality.

Recommended Use Cases

  • Uncensored Content Generation: Ideal for applications requiring minimal content restrictions and direct responses.
  • Creative Writing: Excels in generating diverse and creative text for stories, scenarios, and descriptive passages.
  • Roleplay Scenarios: Highly suitable for interactive roleplaying, offering improved intelligence and reduced repetition compared to its predecessor.

Finetunes
2 models