Geralt-Targaryen commited on
Commit
4a9e505
·
verified ·
1 Parent(s): bf45b22

Update README.md

Browse files
Files changed (1) hide show
  1. README.md +22 -16
README.md CHANGED
@@ -1,16 +1,22 @@
1
- ---
2
- license: odc-by
3
- ---
4
-
5
- Chinese text cleaned from [C4](https://huggingface.co/datasets/allenai/c4) with the following steps:
6
-
7
- - documents containing non-Chinese, non-English text are removed
8
- - documents containing more than 30% English text are removed
9
- - all text in Traditional Chinese is converted into Simplified Chinese using [zhconv](https://github.com/gumblex/zhconv)
10
- - low-quality text (e.g. boilerplates, advertisements) are heuristically removed
11
-
12
- ### statistics
13
-
14
- Number of samples: 32,485,463.
15
-
16
- Size of parquet files: 61G.
 
 
 
 
 
 
 
1
+ ---
2
+ license: odc-by
3
+ ---
4
+
5
+ Chinese text cleaned from [C4](https://huggingface.co/datasets/allenai/c4) with the following steps:
6
+
7
+ - documents containing non-Chinese, non-English text are removed
8
+ - documents containing more than 30% English text are removed
9
+ - all text in Traditional Chinese is converted into Simplified Chinese using [zhconv](https://github.com/gumblex/zhconv)
10
+ - low-quality text (e.g. boilerplates, advertisements) are heuristically removed
11
+
12
+ ### statistics
13
+
14
+ Number of samples: 32,485,463.
15
+
16
+ Size of parquet files: 61G.
17
+
18
+ ### Filtered Version
19
+
20
+ Model-filtered version in the filter branch, including 16,751,263 samples (33G parqeut files).
21
+
22
+ Qwen2.5-32B-Instruct is used to generate language quality annotation (on a scale of 1-5) for 398K Chinese samples and 250K English samples. An XLM-RoBERT-large classifier is trained with regression on these annotations. Any document receiving a score of 1 or 2 from the classifier is removed. The remaining documents are also accompanied by their scores.