Yes, the format is like "conversations": [
{ "from": "user", "value": " what is the color of the dog" }, { "from": "assistant", "value": "it is ...." },
...,
{ "from": "user", "value": " what is the color of the dog" }, { "from": "assistant", "value": "it is ...." },
]
I am interested in understanding whether the current SFT code can support multi-turn dialogues training ?