Spaces:

THUDM

/

GLM-130B

like 530

Running

I was using the default configurations.

Hello, we have deeply analyzed the content generated by the model, it seems that the model does incorrectly split the word into multiple parts in default configurations. Most likely because our Beam Search strategy uses multinomial sampling instead of top-k in order to increase the diversity of the generation, resulting in inaccurate results:

20028 "
117333 凯旋
84066 门
20017 "
115894 的英文
83829 是
20017 "
25754 tri
37796 mph
30232 arch
145656 "。
20052 [
20067 </
20106 s
20033 >
20042 ]

Under BaseStrategy or BeamSearchStrategy (num_beams=4), the output is correct:

Input: "凯旋门"的英文是"[MASK]"
Output: "凯旋门"的英文是"Triumphal Arch "

We will add an option to make beam search use top-k sampling later.

Hi, I have tried again and the result is correct now. Great job! However, it seems it doesn't work well with Traditional Chinese?

1. BeamSearchStrategy
   Input:
   "凱旋門"的英文是"[MASK]"

Output:
"凱旋門"的英文是"" " ""

2. BaseStrategy
   Input:
   "凱旋門"的英文是"[MASK]"

Output:
"凱旋門"的英文是"View Cock ."