Is that correct to do sft without labels, course they mask the prompt sentence like <im_start><im_end>

[lixiangisdemon]

Hello! I truly appreciate the blogs you've posted about LLMs. I've gained a great deal from them. I have a question to ask you. When reading the blog "2024 - 07 - 29_Finetune_Llama31.html", I noticed that when you processed the dataset, you only converted the data into the chat_template format, but there was no process of processing labels. In this case, will the LLM model also learn sentences like <im_start>..<im_end>? And will it reduce the learning accuracy?

[mlabonne]

Hey, thanks! Yes, you want the model to learn the chat template, that's one of the main goals of SFT. The SFTTrainer handles these special tokens appropriately with the dataset_text_field parameter.