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.
