因为原生LLaMA对中文的支持很弱，一个中文汉子往往被切分成多个token，因此需要对其进行中文词表扩展。思路通常是在中文语料库上训练一个中文tokenizer模型，然后将中文tokenizer与LLaMA原生tokenizer进行合并，最终得到一个扩展后的tokenizer模型。国内Chinese-LLaMA-Alpaca开源项目详细说明了词表扩展[2]。
一.对LLaMA tokenizer扩充自定义的词表
原版LLaMA模型的词表大小是32K，其主要针对英语进行训练，下面对其扩充20K中文词表，如下所示：
python merge_tokenizers.py \--llama_tokenizer_dir r'L:/20230902_Llama1/llama-7b-hf' \--chinese_sp_model_file r'./chinese_sp.model'
- llama_tokenizer_dir：指向存放原版LLaMA tokenizer的目录
- chinese_sp_model_file：指向用sentencepiece训练的中文词表文件
说明：在中文通用语料上训练的20K中文词表下载链接参考[3]，如何构建垂直领域的中文词表下次分享。
二.merge_tokenizers.py注释
1.本文环境
本文环境为Windows10，Python3.10，CUDA 11.8，GTX 3090(24G)，内存24G。
2.merge_tokenizers.py代码
import os
from transformers import LlamaTokenizer
from sentencepiece import sentencepiece_model_pb2 as sp_pb2_model
import sentencepiece as spm
import argparse
os.environ["PROTOCOL_BUFFERS_PYTHON_IMPLEMENTATION"] = "python"# parser = argparse.ArgumentParser() # 创建一个ArgumentParser对象
# parser.add_argument('--llama_tokenizer_dir', default=r'L:/20230902_Llama1/llama-7b-hf', type=str, required=True) # 添加参数
# parser.add_argument('--chinese_sp_model_file', default='./chinese_sp.model', type=str) # 添加参数
# args = parser.parse_args() # 解析参数
# llama_tokenizer_dir = args.llama_tokenizer_dir # 这里是LLaMA tokenizer的路径
# chinese_sp_model_file = args.chinese_sp_model_file # 这里是Chinese tokenizer的路径llama_tokenizer_dir = r'L:/20230902_Llama1/llama-7b-hf' # 这里是LLaMA tokenizer的路径
chinese_sp_model_file = r'./chinese_sp.model' # 这里是Chinese tokenizer的路径# 加载tokenizer
llama_tokenizer = LlamaTokenizer.from_pretrained(llama_tokenizer_dir) # 加载LLaMA tokenizer
chinese_sp_model = spm.SentencePieceProcessor() # 定义Chinese tokenizer
chinese_sp_model.Load(chinese_sp_model_file) # 加载Chinese tokenizerllama_spm = sp_pb2_model.ModelProto() # 定义LLaMA tokenizer的sentencepiece model
llama_spm.ParseFromString(llama_tokenizer.sp_model.serialized_model_proto()) # 从LLaMA tokenizer中加载sentencepiece model
chinese_spm = sp_pb2_model.ModelProto() # 定义Chinese tokenizer的sentencepiece model
chinese_spm.ParseFromString(chinese_sp_model.serialized_model_proto()) # 从Chinese tokenizer中加载sentencepiece model# 输出tokens的信息
print(len(llama_tokenizer), len(chinese_sp_model)) # 两个tokenizer的词表大小；输出为32000、20000
print(llama_tokenizer.all_special_tokens) # LLaMA tokenizer的special tokens；输出为['']
print(llama_tokenizer.all_special_ids) # LLaMA tokenizer的special tokens对应的id；输出为[0]
print(llama_tokenizer.special_tokens_map) # LLaMA tokenizer的special tokens；输出为{'bos_token': '', 'eos_token': '', 'unk_token': ''}# 将Chinese tokenizer的词表添加到LLaMA tokenizer中（合并过程）
llama_spm_tokens_set = set(p.piece for p in llama_spm.pieces) # LLaMA tokenizer的词表
print(len(llama_spm_tokens_set)) # LLaMA tokenizer的词表大小；输出为32000
print(f"Before:{len(llama_spm_tokens_set)}") # LLaMA tokenizer的词表大小；输出为Before:32000
for p in chinese_spm.pieces: # 遍历Chinese tokenizer的词表piece = p.piece # Chinese tokenizer的词if piece not in llama_spm_tokens_set: # 如果Chinese tokenizer的词不在LLaMA tokenizer的词表中new_p = sp_pb2_model.ModelProto().SentencePiece() # 创建一个新的sentencepiecenew_p.piece = piece # 设置sentencepiece的词new_p.score = 0 # 设置sentencepiece的scorellama_spm.pieces.append(new_p) # 将sentencepiece添加到LLaMA tokenizer的词表中
print(f"New model pieces: {len(llama_spm.pieces)}") # LLaMA tokenizer的词表大小；输出为New model pieces: 49953# 保存LLaMA tokenizer
output_sp_dir = 'merged_tokenizer_sp' # 这里是保存LLaMA tokenizer的路径
output_hf_dir = 'merged_tokenizer_hf' # 这里是保存Chinese-LLaMA tokenizer的路径
os.makedirs(output_sp_dir, exist_ok=True) # 创建保存LLaMA tokenizer的文件夹
with open(output_sp_dir + '/chinese_llama.model', 'wb') as f:f.write(llama_spm.SerializeToString())
tokenizer = LlamaTokenizer(vocab_file=output_sp_dir + '/chinese_llama.model') # 创建LLaMA tokenizer
tokenizer.save_pretrained(output_hf_dir) # 保存Chinese-LLaMA tokenizer
print(f"Chinese-LLaMA tokenizer has been saved to {output_hf_dir}") # 保存Chinese-LLaMA tokenizer# 测试tokenizer
llama_tokenizer = LlamaTokenizer.from_pretrained(llama_tokenizer_dir) # LLaMA tokenizer
chinese_llama_tokenizer = LlamaTokenizer.from_pretrained(output_hf_dir) # Chinese-LLaMA tokenizer
print(tokenizer.all_special_tokens) # LLaMA tokenizer的special tokens；输出为['<s>', '</s>', '<unk>']
print(tokenizer.all_special_ids) # LLaMA tokenizer的special tokens对应的id；输出为[0, 1, 2]
print(tokenizer.special_tokens_map) # LLaMA tokenizer的special tokens；输出为{'bos_token': '<s>', 'eos_token': '</s>', 'unk_token': '<unk>'}
text = '''白日依山尽，黄河入海流。欲穷千里目，更上一层楼。
The primary use of LLaMA is research on large language models, including'''
print("Test text:\n", text) # 测试文本
print(f"Tokenized by LLaMA tokenizer:{llama_tokenizer.tokenize(text)}") # 测试LLaMA tokenizer
# 输出结果
# Tokenized by LLaMA tokenizer:['▁', '白', '日', '<0xE4>', '<0xBE>', '<0x9D>', '山', '<0xE5>', '<0xB0>', '<0xBD>', '，', '黄', '河', '入', '海', '流', '。', '<0xE6>', '<0xAC>', '<0xB2>', '<0xE7>', '<0xA9>', '<0xB7>', '千', '里', '目', '，', '更', '上', '一', '<0xE5>', '<0xB1>', '<0x82>', '<0xE6>', '<0xA5>', '<0xBC>', '。', '<0x0A>', 'The', '▁primary', '▁use', '▁of', '▁L', 'La', 'MA', '▁is', '▁research', '▁on', '▁large', '▁language', '▁models', ',', '▁including']
print(f"Tokenized by Chinese-LLaMA tokenizer:{chinese_llama_tokenizer.tokenize(text)}") # 测试Chinese-LLaMA tokenizer
# 输出结果
# Tokenized by Chinese-LLaMA tokenizer:['▁白', '日', '依', '山', '尽', '，', '黄河', '入', '海', '流', '。', '欲', '穷', '千里', '目', '，', '更', '上', '一层', '楼', '。', '<0x0A>', 'The', '▁primary', '▁use', '▁of', '▁L', 'La', 'MA', '▁is', '▁research', '▁on', '▁large', '▁language', '▁models', ',', '▁including']
3.生成的目录
参考文献：
[1]是否有基于Llama-2的增量训练模型：https://github.com/ymcui/Chinese-LLaMA-Alpaca/issues/817
[2]https://github.com/ymcui/Chinese-LLaMA-Alpaca/blob/main/scripts/merge_tokenizer/merge_tokenizers.py
[3]https://github.com/ymcui/Chinese-LLaMA-Alpaca/tree/main/scripts/merge_tokenizer/chinese_sp.model
[4]下载Chinese-LLaMA-Alpaca：git clone https://github.com/ymcui/Chinese-LLaMA-Alpaca.git
[5]下载llama-7b-hf：git lfs clone https://huggingface.co/yahma/llama-7b-hf
