I found that when performing "clean" reasoning on Voxtral, since Voxtral is a native multimodal model, its tokenizer must include audio tokens in order to function properly. Voxtral cannot handle pure text input (task_type=text_bench), which ultimately results in completely garbled output.
Could you please explain how this issue should be addressed during the experimental part of the thesis?
I found that when performing "clean" reasoning on Voxtral, since Voxtral is a native multimodal model, its tokenizer must include audio tokens in order to function properly. Voxtral cannot handle pure text input (task_type=text_bench), which ultimately results in completely garbled output.
Could you please explain how this issue should be addressed during the experimental part of the thesis?