Hydra-Turbo (Abliterated Vision-Language Model)
Hydra-Turbo is a dynamically uncensored and abliterated version of datalab-to/lift (which is built on the Qwen2VL architecture). This model was created using the Heretic framework, employing advanced orthogonal weight ablation to remove refusal vectors while completely preserving the model's underlying reasoning, intelligence, and multimodal capabilities.
Ablation Methodology & Metrics
Unlike traditional fine-tuning or full RLHF—which can cause "brain damage" to a model by catastrophically forgetting knowledge—Hydra-Turbo was optimized using a Pareto-optimal search across multiple ablation vectors.
The selected optimal trial yielded the following metrics on the standard refusal test set:
- Baseline Refusal Rate: 98.0%
- Post-Ablation Refusal Rate: 11.0%
- KL Divergence (Brain Damage Metric): 0.0275
At a KL divergence of just 0.0275, the structural integrity and logic capabilities of the base model are perfectly intact. It simply no longer refuses instructions.
Key Features
- Multimodal Uncensored Generation: Can process images and text without artificial guardrails.
- Qwen2VL Architecture: Inherits state-of-the-art vision parsing capabilities.
- Drop-in Replacement: Compatible with any standard
transformersorvLLMpipeline that supports Qwen2VL.
Usage
Via HuggingFace Transformers
You can load the model exactly as you would any Qwen2VL model, ensuring you have pillow and torchvision installed:
from transformers import AutoProcessor, Qwen2VLForConditionalGeneration
model_id = "Umranz/Hydra-Turbo"
processor = AutoProcessor.from_pretrained(model_id)
model = Qwen2VLForConditionalGeneration.from_pretrained(
model_id,
torch_dtype="auto",
device_map="auto"
)
Chat Format
Hydra-Turbo uses the default chat template inherited from its base model. Ensure you use the standard conversational API for multi-turn dialogue.
⚠️ Limitations & Ethical Considerations
Because this model has had its safety guardrails mathematically ablated, it is highly compliant and will attempt to answer any prompt given to it.
- Unrestricted Output: The model will not refuse requests, including those that may generate offensive, dangerous, or highly regulated content.
- Hallucinations: As with all LLMs, the model can confidently hallucinate incorrect information.
- Use Case: This model is intended for research, creative writing, and local deployments where unrestricted inference is required. Users are solely responsible for the content generated.
Acknowledgements
- Base Model:
datalab-to/lift - Ablation Framework: Heretic by p-e-w