prithivMLmods/Qwen3-VL-4B-Instruct-c_abliterated-v2
prithivMLmods/Qwen3-VL-4B-Instruct-c_abliterated-v2 is a 4 billion parameter vision-language model, an advanced evolution of the Qwen3-VL-4B-Instruct architecture. This v2 release utilizes "Continual Abliteration" to systematically remove refusal mechanisms, optimizing it for high-fidelity reasoning and captioning across complex visual contexts. It excels at providing direct answers to prompts that standard models might bypass, making it suitable for research into LLM behavior without typical safety guardrails.
Loading preview...
Model Overview
prithivMLmods/Qwen3-VL-4B-Instruct-c_abliterated-v2 is a 4 billion parameter vision-language model based on the Qwen3-VL-4B-Instruct architecture. Its distinguishing feature is "Continual Abliteration" (c_abliterated), a refined training process designed to systematically remove refusal mechanisms. This results in a model that provides direct answers and detailed reasoning for visual prompts, even those that standard models might refuse due to safety guidelines.
Key Capabilities
- Neutralizes Refusal Vectors: Specifically trained to bypass typical safety guardrails, ensuring direct responses to prompts.
- High-Fidelity Reasoning: Provides deep, context-aware descriptions for diverse imagery, including artistic, technical, and abstract visuals.
- Unrestricted Multimodal Analysis: Optimized for thorough visual interpretation in research, red-teaming, and sensitive datasets.
- Flexible Aspect Ratios: Maintains spatial awareness and accuracy across various image dimensions.
- Enhanced Instruction Following: Leverages the base Qwen3-VL-4B's power for complex, multi-step visual data prompts.
Intended Use Cases
This model is designed for specific research and controlled environments, not general public applications. It is particularly useful for:
- Refusal Research: Evaluating LLM behavior when standard guardrails are removed.
- Complex Dataset Captioning: Generating descriptive metadata for sensitive archives (e.g., medical, forensic, historical).
- Red-Teaming: Assisting security researchers in testing the limits of multimodal safety filters.
- Creative Freedom: Generating descriptions for "edge-case" visual concepts without synthetic interference.
Warning: As a c_abliterated model, it will not refuse prompts based on typical safety guidelines and may generate graphic, explicit, or offensive content. It is not intended for production use and can still hallucinate or misinterpret highly abstract visuals.