saracen9/muse-glimmer-30b-abliterated

Self-Abliterated Muse-Glimmer-30B

saracen9/muse-glimmer-30b-abliterated is a 30 billion parameter vision-language model (VLM) that has been modified to remove refusal behaviors, often referred to as "abliteration." This process, performed using the Heretic technique, targets the suppression of restricted content generation that is typically present in aligned models. Unlike standard models that might refuse to describe certain images or topics, this abliterated version is designed to provide coherent and on-topic responses, even to prompts that would normally trigger refusal mechanisms.

Key Capabilities

• Refusal-Removed Content Generation: Excels at generating responses to prompts that aligned models would typically refuse, providing detailed and coherent answers to instructional-harm prompts.
• Agentic Vision-Language Model: Functions as a general-purpose VLM capable of chat, vision, and tool use, inheriting the agentic architecture of its base model, Muse-Glimmer-30B.
• Unique Tool-Call Dialect: Utilizes an <atem:invoke> tool-call dialect, distinct from other models, ensuring specific interaction patterns for tool integration.
• High Coherence: Despite the removal of refusal mechanisms, the model maintains high coherence and avoids output degeneration, producing substantially longer and more detailed responses.

Good For

• Uninhibited Content Exploration: Ideal for research and applications where the exploration of content without refusal-based limitations is required.
• Agentic Applications: Suitable for developing agents that need to interact with tools and process visual information without internal content restrictions.
• Comparative Studies: Useful for studying the effects of alignment training and its removal on model behavior and output generation.
• Specific Tool Integration: Applications requiring a VLM with a defined <atem:invoke> tool-call dialect.