Datasets:

CASIA-LM
/

ChineseWebText2.0

@@ -39,7 +39,7 @@ We have released the latest and largest Chinese dataset, ChineseWebText 2.0, whi
 We introduce a new toolchain, MDFG-tool (see Figure 1). We begin with the coarse-grained filtering module, which applies rule-based methods to clean the data, focusing on criteria such as text length and sensitive words to ensure data quality. After cleaning, we evaluate the text quality using a BERT-based model. This process generates a quality score, and by selecting an appropriate threshold, we can extract high-quality text data that meets our needs. Next, we use FastText for both single-label and multi-label classification of the cleaned data. Meanwhile, we conduct toxicity assessment. The FastText model is used to filter out toxic content and assign toxicity scores to each text. This scoring system allows researchers to set thresholds for identifying and selecting harmful texts for further training.
-<div align=center><img src="./structure.png" style="zoom:67%;" /></div>
 ### Data Analysis
@@ -50,7 +50,7 @@ In order to provide a high-level overview of the preparation and preprocessing s
 After collecting raw data from various sources, we initially obtain a original Chinese dataset totaling 6.6 TB. However, due to a significant amount of irrelevant and noisy content in some sources, a manual sampling analysis is performed in preparation stage. If irrelevant text accounted for more than 50\% of a source, the data from that source will be discarded entirely. As a result, a substantial portion of the data is removed during the preparation stage, retaining only 67.68\% of the original dataset. In preprocessing stage, four rule-based steps are implemented to filter the remained data. First, the Data Length step remove overly short texts to ensure that each text contains sufficient informational content. Next, the Character Proportion step eliminate texts with a high percentage of noisy characters, such as English, Traditional Chinese characters, or other irrelevant symbols. Finally, the Sensitive Words step and the Deduplication step are employed to remove toxic content and duplicate texts from the dataset. After the preprocessing stage, we produce a high-quality Chinese text dataset totaling 3.8 TB. In the next stage, each text in this high-quality dataset will be enriched with fine-grained annotations, including a quality score, domain lablels, a toxicity score and a toxicity label.
 <div align="center">
-  <img src="/picture/data_statistics.png" width="67%" />
   <br>
     <em>Figure 1: Description of Image 1</em>
 </div>
@@ -62,12 +62,12 @@ After collecting raw data from various sources, we initially obtain a original C
 <table>
   <tr>
     <td>
-    <img src="/picture/quality_distribution.pdf" alt="Image 1" style="width:100%;">
     <br>
     <em>Figure 1: Description of Image 1</em>
     </td>
     <td>
-    <img src="/picture/human_acceptance.pdf" alt="Image 2" style="width:100%;">
     <br>
     <em>Figure 1: Description of Image 1</em>
     </td>

 We introduce a new toolchain, MDFG-tool (see Figure 1). We begin with the coarse-grained filtering module, which applies rule-based methods to clean the data, focusing on criteria such as text length and sensitive words to ensure data quality. After cleaning, we evaluate the text quality using a BERT-based model. This process generates a quality score, and by selecting an appropriate threshold, we can extract high-quality text data that meets our needs. Next, we use FastText for both single-label and multi-label classification of the cleaned data. Meanwhile, we conduct toxicity assessment. The FastText model is used to filter out toxic content and assign toxicity scores to each text. This scoring system allows researchers to set thresholds for identifying and selecting harmful texts for further training.
+<div align=center><img src="./picture/structure.png" style="zoom:67%;" /></div>
 ### Data Analysis
 After collecting raw data from various sources, we initially obtain a original Chinese dataset totaling 6.6 TB. However, due to a significant amount of irrelevant and noisy content in some sources, a manual sampling analysis is performed in preparation stage. If irrelevant text accounted for more than 50\% of a source, the data from that source will be discarded entirely. As a result, a substantial portion of the data is removed during the preparation stage, retaining only 67.68\% of the original dataset. In preprocessing stage, four rule-based steps are implemented to filter the remained data. First, the Data Length step remove overly short texts to ensure that each text contains sufficient informational content. Next, the Character Proportion step eliminate texts with a high percentage of noisy characters, such as English, Traditional Chinese characters, or other irrelevant symbols. Finally, the Sensitive Words step and the Deduplication step are employed to remove toxic content and duplicate texts from the dataset. After the preprocessing stage, we produce a high-quality Chinese text dataset totaling 3.8 TB. In the next stage, each text in this high-quality dataset will be enriched with fine-grained annotations, including a quality score, domain lablels, a toxicity score and a toxicity label.
 <div align="center">
+  <img src="./picture/data_statistics.png" width="67%" />
   <br>
     <em>Figure 1: Description of Image 1</em>
 </div>
 <table>
   <tr>
     <td>
+    <img src="./picture/quality_distribution.pdf" alt="Image 1" style="width:100%;">
     <br>
     <em>Figure 1: Description of Image 1</em>
     </td>
     <td>
+    <img src="./picture/human_acceptance.pdf" alt="Image 2" style="width:100%;">
     <br>
     <em>Figure 1: Description of Image 1</em>
     </td>