I am using a multimodal large model to detect violations of cartoon pornographic images. Because there are many rule classifications (13), but the model's ability to understand the images is not enough, I cannot fine-tune the model, and the prompt has been optimized hundreds of times, but the effect is still not good. What else can I do to improve the model's ability to understand the images and improve the accuracy? I have tried prompt optimization more than 1,000 times, but the effect is still not good. I tried adding a few-shot of various images to help the model understand, but the effect was worse. I tried to increase coT to improve the model's understanding ability, but because the model would have problems with random recognition, the coT was also messy and the effect was not good. I tried to divide it into 4 prompts for partial processing, but the effect was not good because the model's understanding ability was not good. Continue reading...
