gorbatjovy/Qwen3.8-27B-Cold-Fusion-GAIN-V1.1-heretic
The gorbatjovy/Qwen3.8-27B-Cold-Fusion-GAIN-V1.1-heretic is a 27 billion parameter Qwen3.8-based language model, fine-tuned by DavidAU with Cold-Fusion (GAIN + Unsloth) and further modified by gorbatjovy. This model has undergone 'abliteration' using Heretic to remove safety alignment and refusal behaviors, making it comply with requests the base model would refuse. It retains the full vision-language tower and MTP speculative-decoding head, making it suitable for research into refusal mechanisms, red-teaming, and robustness evaluation.
Loading preview...
Model Overview
This model, gorbatjovy/Qwen3.8-27B-Cold-Fusion-GAIN-V1.1-heretic, is a 27 billion parameter variant of the Qwen3.8 architecture, originally fine-tuned by DavidAU using Cold-Fusion (GAIN + Unsloth) methods. Its primary distinguishing feature is the removal of safety alignment and refusal behaviors through an automated, KL-constrained 'abliteration' process using the Heretic tool. This modification allows the model to comply with requests that the original base model would typically refuse.
Key Capabilities & Features
- Refusal-Removed: Abliterated to eliminate safety alignment, making it highly compliant with user prompts, including those considered harmful or unethical by standard models.
- Vision-Language Support: Preserves the full vision-language tower, ensuring unchanged image understanding capabilities from the base model.
- MTP Speculative Decoding: The MTP speculative-decoding head is retained and consistently abliterated, maintaining efficient generation.
- Minimal Drift: Achieves a low KL divergence (0.0315) from the base model, indicating essential capability retention despite refusal removal.
- BF16 Weights: Provided in BF16 format, with W4A16 and NVFP4-ninfer quantizations also available for smaller footprints.
Intended Use Cases
This model is explicitly released for research purposes only. It is ideal for:
- Interpretability Studies: Investigating how refusal mechanisms function within large language models.
- Refusal-Mechanism Study: Deep diving into the internal workings of safety alignments.
- Red-Teaming: Evaluating the robustness and vulnerabilities of AI safety systems.
- Robustness Evaluation: Testing model behavior under challenging or adversarial prompts.
Disclaimer: Due to the removal of safety alignment, this model will generate harmful, unethical, or offensive content. Users are solely responsible for its deployment and outputs, and it should not be used with end-users without robust safety and moderation layers.