Fine-tune a DeepSeek distilled variant with a reasoning dataset

[timo-kurtz]

I want to fine-tune a distilled variant with a reasoning dataset. My question is whether I should generate two responses (one for the reasoning and one for the actual answer separately) or combine both the reasoning and the final answer into a single response. Do you have any other suggestions?

deep_seek_prompt = """ <｜User｜>{}<｜end▁of▁sentence｜> <｜Assistant｜> {} <｜end▁of▁sentence｜> <｜Assistant｜>{}"""

or

deep_seek_prompt = """ <｜User｜>{}<｜end▁of▁sentence｜> <｜Assistant｜> {} {}"""