commit fca1c541bac3c8dba5c763532863d73a188f8849
Author: ViperEkura <3081035982@qq.com>
Date:   Mon Jun 30 12:26:37 2025 +0800

    提交数据下载代码

diff --git a/.gitignore b/.gitignore
new file mode 100644
index 0000000..b2426b8
--- /dev/null
+++ b/.gitignore
@@ -0,0 +1,8 @@
+# cache
+__pycache__*
+
+# dataset
+dataset/*
+
+# tokenzier
+tokenizer.json
diff --git a/chinese-c4.py b/chinese-c4.py
new file mode 100644
index 0000000..f0934f6
--- /dev/null
+++ b/chinese-c4.py
@@ -0,0 +1,28 @@
+from datasets import load_dataset
+import json
+import os
+
+if __name__ == "__main__":
+    dataset_dict = load_dataset("shjwudp/chinese-c4")
+    train_dataset = dataset_dict["train"]
+
+    chunk_size = 1000000
+    total_samples = len(train_dataset)
+    num_chunks = (total_samples // chunk_size) + 1
+    script_dir = os.path.dirname(os.path.abspath(__file__))
+    output_dir = os.path(script_dir, "dataset", "chinese-c4")
+    os.makedirs(output_dir, exist_ok=True)
+
+    for i in range(num_chunks):
+        start_idx = i * chunk_size
+        end_idx = min((i + 1) * chunk_size, total_samples)
+        chunk = train_dataset.select(range(start_idx, end_idx))
+
+        output_path = f"{output_dir}/{output_dir}_text_chunk_{i}.jsonl"
+        with open(output_path, "w", encoding="utf-8") as f:
+            for example in chunk:
+                # 每行写入一个 {"text": "xxx"} 对象
+                json_line = {"text": example["text"]}
+                f.write(json.dumps(json_line, ensure_ascii=False) + "\n")
+
+        print(f"Saved text chunk {i} to {output_path}")
diff --git a/english-fineweb.py b/english-fineweb.py
new file mode 100644
index 0000000..ecf929c
--- /dev/null
+++ b/english-fineweb.py
@@ -0,0 +1,42 @@
+from datasets import load_dataset
+import json
+import os
+import re
+
+def comprehensive_normalization(text):
+    replacements = {
+        '\u2018': "'", '\u2019': "'", '\u0060': "'",
+        '\u201C': '"', '\u201D': '"', 
+        '\u2013': '-', '\u2014': '--', '\u2212': '-',
+        '\u00A0': ' ',
+        '\u2026': '...'
+    }
+    pattern = re.compile('|'.join(re.escape(k) for k in replacements))
+    return pattern.sub(lambda m: replacements[m.group()], text)
+
+if __name__ == "__main__":
+    dataset_dict = load_dataset("HuggingFaceFW/fineweb","sample-10BT")
+    train_dataset = dataset_dict["train"]
+
+    chunk_size = 1000000
+    total_samples = len(train_dataset)
+    num_chunks = (total_samples // chunk_size) + 1
+
+    script_dir = os.path.dirname(os.path.abspath(__file__))
+    output_dir = os.path(script_dir, "dataset", "english-fineweb")
+    os.makedirs(output_dir, exist_ok=True)
+
+    for i in range(num_chunks):
+        if i == 10: 
+            break
+        start_idx = i * chunk_size
+        end_idx = min((i + 1) * chunk_size, total_samples)
+        chunk = train_dataset.select(range(start_idx, end_idx))
+        
+        output_path = f"{output_dir}/{output_dir}text_chunk_{i}.jsonl"
+        with open(output_path, "w", encoding="utf-8") as f:
+            for example in chunk:
+                json_line = {"text": comprehensive_normalization(example["text"])}
+                f.write(json.dumps(json_line, ensure_ascii=False) + "\n")
+        
+        print(f"Saved text chunk {i} to {output_path}")
\ No newline at end of file
diff --git a/to_ids.py b/to_ids.py
new file mode 100644
index 0000000..1726624
--- /dev/null
+++ b/to_ids.py
@@ -0,0 +1,5 @@
+from tokenizer import BpeTokenizer
+
+
+if __name__ == "__main__":
+    tokenzier = BpeTokenizer("tokenizer.json")
diff --git a/tokenizer.py b/tokenizer.py
new file mode 100644
index 0000000..689e3fc
--- /dev/null
+++ b/tokenizer.py
@@ -0,0 +1,119 @@
+from tokenizers import Tokenizer, Encoding
+from tokenizers import decoders, processors, normalizers, pre_tokenizers
+from tokenizers.models import BPE
+from tokenizers.trainers import BpeTrainer
+
+from typing import List, Union
+import concurrent.futures
+
+class BpeTokenizer:
+    def __init__(self, path=None):
+        self._control_tokens = ["<bos>", "<eos>", "<pad>"]
+        self._special_tokens = ["<|user|>", "<|system|>"]
+        model = BPE()
+        tokenizer = Tokenizer(model)
+        tokenizer.normalizer = normalizers.Sequence([
+            normalizers.NFC()
+        ])
+        tokenizer.pre_tokenizer = pre_tokenizers.Sequence([
+            pre_tokenizers.Punctuation(behavior="isolated"),
+            pre_tokenizers.Metaspace(prepend_scheme="never"),
+            pre_tokenizers.Split(pattern=r"(\d+|[a-zA-Z]+|(?:'s|'t|'re|'ve|'m|'ll|'d))", behavior="isolated"),
+            pre_tokenizers.ByteLevel(add_prefix_space=False, use_regex=False)
+        ])
+        tokenizer.decoder = decoders.Sequence([
+            decoders.ByteLevel(),
+            decoders.Metaspace(prepend_scheme="never")
+        ])
+        tokenizer.post_processor = processors.Sequence([
+            processors.ByteLevel(trim_offsets=False)
+        ])
+        self._tokenizer = tokenizer
+        
+        if path is not None:
+            self._tokenizer = Tokenizer.from_file(path)
+    
+    def __init_trainer(self, vocab_size, min_freq):
+        alphabet = pre_tokenizers.ByteLevel.alphabet()
+        min_size  = len(alphabet) + len(self._control_tokens)
+        assert vocab_size > min_size
+        
+        trainer = BpeTrainer(
+            vocab_size=vocab_size,
+            min_frequency=min_freq, 
+            limit_alphabet= vocab_size // 4,
+            max_token_length=18,
+            special_tokens=self._control_tokens,
+            show_progress=True,
+            initial_alphabet=alphabet,
+        )
+        return trainer
+        
+    def _prepare_trainer_and_tokens(self, vocab_size: int, min_freq: int, reserved_token_size: int) -> tuple:
+        assert reserved_token_size > len(self._special_tokens)
+        reserved_tokens = [f"<|rsv{i:02d}|>" for i in range(reserved_token_size - len(self._special_tokens))]
+        detail_vocab_size = vocab_size - (len(reserved_tokens) + len(self._special_tokens))
+        trainer = self.__init_trainer(docab_size=detail_vocab_size, min_freq=min_freq)
+        return trainer, detail_vocab_size, reserved_tokens
+
+    def train(self, files, vocab_size, min_freq, reserved_token_size=100):
+        trainer, _, reserved_tokens = self._prepare_trainer_and_tokens(
+            vocab_size=vocab_size,
+            min_freq=min_freq,
+            reserved_token_size=reserved_token_size
+        )
+        self._tokenizer.train(files=files, trainer=trainer)
+        self._tokenizer.add_special_tokens(self._special_tokens + reserved_tokens)
+            
+    def train_from_iterator(self, iterator, vocab_size, min_freq, reserved_token_size=100):
+        trainer, _, reserved_tokens = self._prepare_trainer_and_tokens(
+            vocab_size=vocab_size,
+            min_freq=min_freq,
+            reserved_token_size=reserved_token_size
+        )
+        self._tokenizer.train_from_iterator(iterator=iterator, trainer=trainer)
+        self._tokenizer.add_special_tokens(self._special_tokens + reserved_tokens)
+            
+    def save(self, path):
+        self._tokenizer.save(path)
+        
+    def load(self, path):
+        self._tokenizer = Tokenizer.from_file(path)
+
+    def encode(self, tokens: Union[str, List[str]], out_ids=True, num_threads=4) -> List:
+        if isinstance(tokens, str):
+            encoded: Encoding = self._tokenizer.encode(tokens)
+            return encoded.ids if out_ids else encoded.tokens
+        else:
+            with concurrent.futures.ThreadPoolExecutor(max_workers=num_threads) as executor:
+                encodings: List[Encoding] = list(executor.map(self._tokenizer.encode, tokens))
+            
+            if out_ids:
+                return [encoding.ids for encoding in encodings]
+            else:
+                return [encoding.tokens for encoding in encodings]
+
+    def decode(self, tokens: List[int]) -> str:
+        return self._tokenizer.decode(tokens)
+    
+    def __len__(self) -> int:
+        return self._tokenizer.get_vocab_size()
+    
+    @property
+    def stop_ids(self) -> List[int]:
+        stop_ids = []
+        for token in self._control_tokens:
+            stop_ids.append(self._tokenizer.token_to_id(token))
+        return stop_ids
+    
+    @property
+    def bos_id(self) -> int:
+        return self._tokenizer.token_to_id("<bos>")
+    
+    @property
+    def eos_id(self) -> int:
+        return self._tokenizer.token_to_id("<eos>")
+    
+    @property
+    def pad_id(self) -> int:
+        return self._tokenizer.token_to_id("<pad>")