Model size does not match the article

[smurf-1119]

I found that the model you used in the script is opt-350m, while the one mentioned in the article is opt-125m, indicating different versions. Would this affect the final results? Additionally, I noticed that the model training did not employ reference summary loss training. Would this affect the final results? And, I found multiple errors in the code under raw mode. Are there any plans to fix them in the near future?