notes/sft.txt

假如原始数据如下：

文本部分：38为eot,数据类型为list of list
[
    [1, 2, 3, 38],
    [4, 5, 38],
    [6, 7, 8, 9, 38]
]
音频部分：
[
    [100,101,102],
    [103,104,105],
    [106,107,108,109,110],
]

处理出以下内容

text_input_ids,用38作为padding:
[
    [1, 2, 3, 38, 38],
    [4, 5, 38, 38, 38],
    [6, 7, 8, 9, 38]
]
text_attention_mask:
[
    [1,1,1,1,0],
    [1,1,1,0,0],
    [1,1,1,1,1]
]
audio_input_ids,每个数据后面添加一个8192，然后用8192作为padding:
[
    [100,101,102,8192,8192,8192],
    [103,104,105,8192,8192,8192],
    [106,107,108,109,110，8192],
]

target:
[
    [0, 0, 0, 100, 101, 102, 8193, 8193, 8193, 8193],
    [0, 0, 103, 104, 105, 8193, 8193, 8193, 8193, 8193],
    [0, 0, 0, 0, 106, 107, 108, 109, 110, 8193]
]

loss-mask:
[
    [0,0,0,1,1,1,1,0,0,0],
    [0,0,1,1,1,0,0,0,0,0],
    [0,0,0,0,1,1,1,1,1,0]
]


模型的输入(模型内部处理)：
[
    [1, 2, 3, 0, 100, 101, 102, 8193, 8193, 8193],
    [4, 5, 0, 103, 104, 105, 8193, 8193, 8193, 8193],
    [6, 7, 8, 9, 0, 106, 107, 108, 109, 110]

]