Do we have finetuned models for more accurate icon classification?

[tangkangqi]

The icon detection is supper great, however the icon classification ( parsed_content_list 'content' field is not so accurate ) If more accurate icon detection result we can feed to LLM directly to generate actions.