saracen9/vt-vlm-base-qwen35

Overview

The saracen9/vt-vlm-base-qwen35 is a 9-billion parameter vision-language model derived from Qwen/Qwen3.5-9B. Its core distinction lies in the surgical removal of refusal and evasive-hedging behaviors through a process called "abliteration" using the Heretic tool. This modification ensures the model provides plain, literal descriptions of visual content, including potentially sensitive imagery, without deflection or moralization.

Key Capabilities & Features

• Refusal-Removed Vision-Language Model: Specifically engineered to describe visual content directly, without the inherent biases or refusal mechanisms often found in RLHF-trained models.
• Base for VLM Fine-tuning: Designed as a foundational model for the video-trainer pipeline, supporting the training of small, cost-effective LoRA adapters for specialized captioning tasks.
• Flexible Deployment: Provided as a full HF-transformers-format checkpoint (safetensors) for fine-tuning, and as a two-file GGUF (Q4_K_M-quantized text model and F16 vision projector) for local inference with llama-server.
• Optimized for llama-server: Includes specific instructions for local serving with llama-server to manage its 32768-token context size efficiently, avoiding OOM errors.

Good For

• Literal Image/Video Captioning: Ideal for applications requiring unbiased, factual descriptions of visual data, regardless of content sensitivity.
• Developing Specialized VLMs: Serves as an excellent base for researchers and developers looking to fine-tune vision-language models with custom LoRA adapters for niche captioning concepts.
• Local Inference: The provided GGUF format allows for efficient local deployment and inference, particularly with llama-server's router mode for managing multiple adapters.