---

base_model:

• prithivMLmods/Qwen3-VL-8B-Instruct-abliterated-v2
• Qwen/Qwen3-VL-8B-Instruct

---

NSFWVision qwen3 vl 8b V3 (GGUF)

GGUF quants for https://huggingface.co/GitMylo/nsfwvision-qwen3-vl-8b-v3-safetensors. An experimental model with the vision tower and base model trained, the mmproj (translator) frozen. Afterwards, the trained base model was reset, acting like a sponge during training.

A finetune of the vision tower of Qwen3-VL-8B-Instruct-abliterated-v2.
During training, the model was unfrozen, after training finished, the model was reset to the original weights (abliterated). Allowing the model to train and later resetting it allows it to "soak" in the training and avoid biasing the model's style of writing and prompt following capabilities. It was used effectively as a sponge to catch loss related to sentence structure, allowing the vision tower to actually learn the concepts.