You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
Currently, the PIPELINE class in src/util.py has a arg "stop_token" which means the special designed single token_id to stop generation.
But in most cases, the stop_token should be a token id list. For an example, if the prompt looks like :
"User:请根据以下材料设计一道中餐菜谱。要求生成菜名和具体做法,菜谱最后以”完成!“结束。材料:猪后腿肉,青椒,洋葱,盐,胡椒。\nAssistant:菜名:"
The results should looks like below:
Currently, the PIPELINE class in src/util.py has a arg "stop_token" which means the special designed single token_id to stop generation.
But in most cases, the stop_token should be a token id list. For an example, if the prompt looks like :
"User:请根据以下材料设计一道中餐菜谱。要求生成菜名和具体做法,菜谱最后以”完成!“结束。材料:猪后腿肉,青椒,洋葱,盐,胡椒。\nAssistant:菜名:"
The results should looks like below:
The stop_token should be set like below:
end_token = pipeline.encode("完成!")
In current implementation, the end_token is not able to stop generation.
I just made an update in my fork to supply the stop_words implementation.
The text was updated successfully, but these errors were encountered: