Update README.md

README.md

@@ -17,4 +17,82 @@ configs:
 ---
 # Dataset Card for "vietnamese-mlmcorpus"
 
+## How can we setup build dataset 
+Prepare several functions: split puntual, tokenizer, and use HF dataset for optimize processing
+```
+def get_tokens(examples):
+    '''
+      Tokenizer samples into token_ids
+    '''
+    return tokenizer(examples)['input_ids']
+
+def truncation(passage, pattern= '[.\n]'):
+    '''
+      Pattern split passage
+    '''
+    output = re.split(pattern, passage)
+    output = [item for item in output if len(item.split()) > 0]
+
+    return output
+
+def split_puntual(example, threshold=512):
+    '''
+      Split a long documents into spans with ~512 tokens 
+    '''
+    texts = truncation(example)
+    tokenized = get_tokens(texts)
+
+    tmp, group = [], []
+    count = 0
+    for tokens, text in zip(tokenized, texts):
+        count += len(tokens)
+        if count <= threshold:
+            tmp.append(text.strip())
+        else:
+            if len(tmp) > 0:
+                group.append('. '.join(tmp)) # update List[str]
+                count = len(tokens) # set count at current idx
+                tmp = [] # reset list
+                tmp.append(text.strip())
+            else:
+                count = 0
+                group.append('summary')
+    return group
+
+def process(examples):
+    '''
+      Build a stack processing
+    '''
+    chunks = []
+    for x in examples:
+        chunks += split_puntual(x)
+    return {'text':chunks}
+```
+
+Now, we run with this code
+```
+import re
+from datasets import load_dataset
+from transformers import AutoTokenizer
+
+tokenizer = AutoTokenizer.from_pretrained('google/mt5-small')
+
+if __name__ == '__main__':
+    dataset = load_dataset("ademax/binhvq-news-corpus")
+    print("Total original: ", dataset)
+    
+    dataset = dataset.map(
+       lambda example : process(example['content']),
+       num_proc=2, batched=True,
+       remove_columns=['content', 'title', 'summary', 'category']
+       )
+
+    # filter samples less than 30 words
+    dataset = dataset.filter(lambda example: len(example['text'].split(' ')) > 30)
+    print("Processing: ", dataset)
+
+    dataset = dataset.train_test_split(test_size=0.0002)
+    dataset.save_to_disk('release')
+```
+
 [More Information needed](https://github.com/huggingface/datasets/blob/main/CONTRIBUTING.md#how-to-contribute-to-the-dataset-cards)