Qwen

/

Qwen2.5-1.5B-Instruct

like 381

Follow

Qwen 23k

Why tokenizer.all_special_tokens has vision-related special tokens?
This is not a multimodal model, right?
Were these tokens used in any way when training the model?

['<|im_end|>',
 '<|endoftext|>',
 '<|im_start|>',
 '<|object_ref_start|>',
 '<|object_ref_end|>',
 '<|box_start|>',
 '<|box_end|>',
 '<|quad_start|>',
 '<|quad_end|>',
 '<|vision_start|>',
 '<|vision_end|>',
 '<|vision_pad|>',
 '<|image_pad|>',
 '<|video_pad|>']