Results are not reproducible on Qwen1.5-1.8B

[Fantasy1120]

Hi, I tried training with Qwen1.5-1.8, and I found that the results varied greatly each time.
For example, I trained three times and the corresponding evaluations are 46,61,55 (GQA); 37,53,43 (TextVQA).
I have followed your default training settings, i.e, global batch size, lr and conv_version, and I'd like to know what caused such a big difference?

Additionally, I would like to ask if I want to add a new LLM, then how can I find the template for this LLM? For example the template for Qwen2.

Thanks for your answer.