File size: 2,329 Bytes
560b763
 
 
 
 
 
 
 
0aa9c28
 
 
560b763
 
3dc9a1c
 
560b763
 
 
 
 
 
 
 
6f5ca5b
 
560b763
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
---
language:
- ja
- zh
- en
pipeline_tag: text-generation
---
[ChatGLM3-6B](https://github.com/THUDM/ChatGLM3)是一个中英双语大模型,本项目为ChatGLM3-6B加入日文能力。  
通过sentencepiece在日文wiki和青空文库上训练BPE的tokenizer,词表扩充14000(64789 -> 78554),相比原生ChatGLM3的日文编码效率提升接近一倍,中文英文编码效率不变。  

扩充词表后,在1B tokens上进行增量预训练,然后在22万对话上进行指令微调,得到最终的模型。

共有两个模型:
- [ChatGLM3-Japanese-Zero]( https://huggingface.co/dummy-foo/ChatGLM3-Japanese-Zero ):经过扩词表和resize后的模型,保留了ChatGLM3的中英文能力,尚无日文能力,但因为编码效率高,适合在日文语料上训练。
- [ChatGLM3-Japanese](https://huggingface.co/dummy-foo/ChatGLM3-Japanese):本模型,对ChatGLM3-Japanese-Zero进行日文语料增量预训练和指令微调的模型,可以日文对话。

[GitHub](https://github.com/akiragy/ChatGLM3-Japanese)上有全流程的代码,包含训练tokenizer和训练模型。

使用样例:
```python 
import torch
from transformers import AutoModelForCausalLM, AutoTokenizer

tokenizer = AutoTokenizer.from_pretrained("dummy-foo/ChatGLM3-Japanese", trust_remote_code=True)
model = AutoModelForCausalLM.from_pretrained("dummy-foo/ChatGLM3-Japanese", trust_remote_code=True, device_map='cuda:0', torch_dtype=torch.float16)

# test Chinese chat
response, _ = model.chat(tokenizer, "你是谁", history=[])
print("\nQ:你是谁,A:", response)
response, _ = model.chat(tokenizer, "你喜欢甜粽子还是咸粽子", history=[])
print("\nQ:你喜欢甜粽子还是咸粽子,A:", response)

# test Japanese chat
response, _ = model.chat(tokenizer, "あなたは誰ですか", history=[])
print("\nQ:あなたは誰ですか,A:", response)
response, _ = model.chat(tokenizer, "すみません、ちょっとお聞きしたいことがあるんですが", history=[])
print("\nQ:すみません、ちょっとお聞きしたいことがあるんですが,A:", response)

# test Japanese tokenizer
print("\ntokens: ", tokenizer.tokenize("恥の多い生涯を送って来ました。自分には、人間の生活というものが、見当つかないのです。"))
```