[minillm] what is the PROMPT_DATA_DIR and LM_DATA_DIR

[Harryjun]

[minillm] what is the PROMPT_DATA_DIR and LM_DATA_DIR ? LM_DATA_DIR is required?

[t1101675]

PROMPT_DATA_DIR contains the prompts in $\mathcal{D}$, the conditional generation dataset.
LM_DATA_DIR contains the data in $\mathcal{D}_{\text{PT}}$, the long-document texts.
Details of $\mathcal{D}$ and $\mathcal{D}_{\text{PT}}$ can be found in our paper

[Harryjun]

@t1101675 展开说说？我这里只有一份SFT数据，训了一个14b和3b想把14b效果蒸馏到3b可行吧？
SFT数据我理解就是PROMPT_DATA_DIR？

[t1101675]

可行。可以参考此文件运行。