可以公布以下具体训练细节吗？

[bingwork]

您好！ 有些训练的具体细节，看了两三遍论文，还是没有理解，还望可行的话，能够说的具体一些。 1，请问具体用了什么数据训练？是只用了ChartLlama dataset吗？ 2，请问训练数据具体训练了哪个阶段？只是做了stage 2的visual instruction tuning ，基于lora的训练吗？还是stage 1 也做了训练？ 3，请问训练epoch是多少？是一个，还是三个呢？ 多谢！ @tingxueronghua

[tingxueronghua]

你好！

1. 我们只用了ChartLlama
2. 可以参考llava的官方实现：https://github.com/haotian-liu/LLaVA/blob/main/scripts/v1_5/finetune_task_lora.sh 只不过我们在训练的时候，llava-1.5官方还没有支持类似的功能，所以是自己实现的，事后我根据git commit的记录确认了我们的修改是全面且正确的。
3. 3个epoch。不过最早的时候llava的代码里似乎有一点小问题，按照最新的脚本大概训一个epoch就够了。

Hello!

1. We only used ChartLlama.
2. You can refer to the official implementation of Llava: https://github.com/haotian-liu/LLaVA/blob/main/scripts/v1_5/finetune_task_lora.sh. However, when we were training, Llava-1.5 officially did not support similar functions, so we implemented it ourselves. Afterward, I confirmed through the git commit records that our modifications were comprehensive and correct.
3. 3 epochs. However, there seemed to be a small issue in the Llava code in the early stages. According to the latest script, training for about one epoch should be enough.