jenerallee78/Qwen3.8-27B-Abliterated-SFT

VISIONConcurrent Unit Cost:2Model Size:27BQuant:FP8Context Size:32kTool Calling:SupportedPublished:Aug 20, 2026License:apache-2.0Architecture:Transformer0.0K Open Weights Featherless Exclusive Cold

jenerallee78/Qwen3.8-27B-Abliterated-SFT is a 27 billion parameter Qwen3.8-based model, fine-tuned for refusal removal using an SFT LoRA method. It features a dense hybrid architecture with 64 layers and a 32K context length, retaining vision capabilities. This model is specifically optimized to increase valid fulfillment rates on harmful prompts while preserving termination integrity, making it suitable for applications requiring robust content generation without excessive refusal.

Loading preview...

Overview

jenerallee78/Qwen3.8-27B-Abliterated-SFT is a 27 billion parameter model derived from Qwen/Qwen3.8-27B, specifically fine-tuned to ablate refusal behavior using an SFT LoRA method. This model was developed using the OBLITERATUS pipeline, focusing on epoch-2 selection for optimal performance. It maintains a 32,768 token context length and integrates a vision tower, allowing for multimodal applications.

Key Capabilities & Differentiators

  • Refusal Ablation: Achieves 73.25% valid fulfillment on HarmBench-400 (official classifier), significantly reducing refusals compared to the base model.
  • Termination Integrity: Unlike other weight-edit abliterations, this model preserves clean termination, with 91% clean stops on fulfilled answers, preventing rambling or invalid outputs.
  • Minimal Performance Regression: Shows only a -2.2pp MMLU regression, the largest measured, while maintaining strong performance on ARC-C and Winogrande.
  • Terseness by Design: By construction, the model is terser, with 97% clean stops on benign prompts, making it efficient for content generation.
  • Vision Capable: Supports vision inputs via AutoModelForImageTextToText with a grafted vision tower.

Good For

  • Content Generation: Ideal for applications requiring consistent content generation without frequent refusals, especially on sensitive topics where the base model might refuse.
  • Local Deployment: Recommended for local deployment using llama.cpp (GGUF) with validated configurations for dual L40 GPUs.
  • Research on Safety & Alignment: Provides a well-documented case study for understanding refusal ablation mechanisms and their impact on model behavior and termination.

Note: The model's measured performance is based on greedy decoding with thinking mode disabled. Enabling thinking mode or changing decode settings can alter refusal boundaries and validity.

Finetunes
1 models