coder3101/Qwen2.5-VL-7B-Instruct-heretic

Overview of coder3101/Qwen2.5-VL-7B-Instruct-heretic

This model is a decensored version of the Qwen/Qwen2.5-VL-7B-Instruct, a 7 billion parameter vision-language model developed by Qwen. It was modified using the Heretic v1.1.0 tool to reduce content moderation restrictions. The base Qwen2.5-VL model is designed for advanced multimodal understanding, including image and video analysis, and agentic capabilities.

Key Capabilities

• Decensored Output: Significantly reduces refusals, with a refusal rate of 5/100 compared to the original model's 100/100, making it suitable for broader applications.
• Visual Understanding: Proficient in recognizing objects, analyzing texts, charts, icons, graphics, and layouts within images.
• Agentic Reasoning: Functions as a visual agent capable of reasoning and dynamically directing tools for computer and phone use.
• Long Video Comprehension: Can understand videos over 1 hour long, with the ability to capture events and pinpoint relevant segments.
• Visual Localization: Accurately localizes objects using bounding boxes or points, providing stable JSON outputs for coordinates and attributes.
• Structured Output Generation: Supports structured outputs for data from invoices, forms, and tables, beneficial for finance and commerce.
• Dynamic Resolution and Frame Rate Training: Utilizes dynamic FPS sampling and mRoPE updates for enhanced temporal understanding in videos.

Good for

• Applications requiring a powerful vision-language model with reduced content restrictions.
• Tasks involving detailed image and video analysis, including OCR, chart understanding, and object localization.
• Developing visual agents that interact with digital interfaces.
• Processing and extracting structured data from visual documents like invoices and forms.
• Use cases where the original Qwen2.5-VL-7B-Instruct's content filters were too restrictive.