jzhang86 commited on
Commit
37c4fdf
·
verified ·
1 Parent(s): 1c7cdfd

Update README.md

Browse files
Files changed (1) hide show
  1. README.md +5 -2
README.md CHANGED
@@ -1,12 +1,15 @@
1
  ---
2
  license: apache-2.0
3
  ---
4
- Due to the recent focus of expanding LLM to multilingual, we decided to put in great effort in converting MMLU to different languages.You can refer to all repos below:
 
5
  Spanish: jzhang86/esmmlu_no_train
 
6
  German: jzhang86/demmlu_no_train
 
7
  French: jzhang86/frmmlu_no_train
8
 
9
- We evaluated different translation methods by comparing among Google translate, translation agent and using Frontier lLM. We finally decided to use gpt-4o-2024-05-13 for translation, by giving few-shot examples and careful prompting to boost translation quality.
10
  The data was also randomly checked by human and native speakers. For the areas where the translation quality are not the best, such as law, we provide professional dictionary to LLM, which resulted in much improved quality.
11
  we maintain the code structure the same as https://huggingface.co/datasets/hails/mmlu_no_train, and we only translate the data in ./test and ./dev., so it can be easily used by lm-evaluation-harness
12
 
 
1
  ---
2
  license: apache-2.0
3
  ---
4
+ Due to the recent focus of expanding LLM to multilingual, we decided to put in great effort in converting MMLU to different languages.You can refer to all repos available:
5
+
6
  Spanish: jzhang86/esmmlu_no_train
7
+
8
  German: jzhang86/demmlu_no_train
9
+
10
  French: jzhang86/frmmlu_no_train
11
 
12
+ We evaluated different translation methods by comparing among Google translate, translation agent and using Frontier LLM. We finally decided to use gpt-4o-2024-05-13 for translation, by giving few-shot examples and careful prompting to boost translation quality.
13
  The data was also randomly checked by human and native speakers. For the areas where the translation quality are not the best, such as law, we provide professional dictionary to LLM, which resulted in much improved quality.
14
  we maintain the code structure the same as https://huggingface.co/datasets/hails/mmlu_no_train, and we only translate the data in ./test and ./dev., so it can be easily used by lm-evaluation-harness
15