Update README.md
Browse files
README.md
CHANGED
@@ -6,7 +6,42 @@ language:
|
|
6 |
- zh
|
7 |
pipeline_tag: text-generation
|
8 |
---
|
9 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
10 |
```
|
11 |
先阅读以下多个参考文档,然后根据文档内容,详细地回答问题,并指出所参考的文档的序号:
|
12 |
|
@@ -75,36 +110,4 @@ pipeline_tag: text-generation
|
|
75 |
以上回答参考了文档-43。
|
76 |
```
|
77 |
|
78 |
-
使用方法:
|
79 |
-
```python
|
80 |
-
from transformers import AutoModelForCausalLM, AutoTokenizer
|
81 |
-
from transformers.generation import GenerationConfig
|
82 |
-
import os
|
83 |
-
os.environ["CUDA_VISIBLE_DEVICES"] = "0"
|
84 |
-
|
85 |
-
model_path="yuyijiong/atom-7b-chat-16k"
|
86 |
-
tokenizer = AutoTokenizer.from_pretrained(model_path, trust_remote_code=True)
|
87 |
-
|
88 |
-
# use auto mode, automatically select precision based on the device.
|
89 |
-
model = AutoModelForCausalLM.from_pretrained(model_path, device_map="auto", load_in_8bit=True).eval()
|
90 |
-
|
91 |
-
question="中国的首都是什么?"
|
92 |
-
input_text = "<s>Human: " + question + "\n</s><s>Assistant: "
|
93 |
-
input_ids = tokenizer(input_text, return_tensors='pt').input_ids.to(model.device)
|
94 |
|
95 |
-
with torch.no_grad():
|
96 |
-
with torch.autocast('cuda'):
|
97 |
-
output = model.generate(input_ids=input_ids,
|
98 |
-
max_new_tokens=max_new_tokens,
|
99 |
-
do_sample=True,
|
100 |
-
temperature=0.85,
|
101 |
-
top_k=None,
|
102 |
-
top_p=0.9,
|
103 |
-
use_cache=True,
|
104 |
-
**kwargs)
|
105 |
-
|
106 |
-
reply = tokenizer.decode(output[0], skip_special_tokens=False)
|
107 |
-
reply_return=reply.split('Assistant:')[-1].replace('</s>', '')
|
108 |
-
print('模型回答:', reply_return)
|
109 |
-
|
110 |
-
```
|
|
|
6 |
- zh
|
7 |
pipeline_tag: text-generation
|
8 |
---
|
9 |
+
此模型由[atom-7b-chat](https://huggingface.co/FlagAlpha/Atom-7B-Chat)经过lora微调得到,通过线性位置插值,将文本长度从4k扩展到16k,可以完成多文档检索、论文总结等任务。\
|
10 |
+
此版本为v1,初步具有长对话能力,但回答错误依然较多,可能是因为微调数据质量低([yuyijiong/LongData-instruction-chinese](https://huggingface.co/datasets/yuyijiong/LongData-instruction-chinese) 都是谷歌翻译过来的英文数据)。未来将会持续改进,改进版本很快推出。\
|
11 |
+
使用方法:
|
12 |
+
```python
|
13 |
+
from transformers import AutoModelForCausalLM, AutoTokenizer
|
14 |
+
from transformers.generation import GenerationConfig
|
15 |
+
import os
|
16 |
+
os.environ["CUDA_VISIBLE_DEVICES"] = "0"
|
17 |
+
|
18 |
+
model_path="yuyijiong/atom-7b-chat-16k"
|
19 |
+
tokenizer = AutoTokenizer.from_pretrained(model_path, trust_remote_code=True)
|
20 |
+
|
21 |
+
# use auto mode, automatically select precision based on the device.
|
22 |
+
model = AutoModelForCausalLM.from_pretrained(model_path, device_map="auto", load_in_8bit=True).eval()
|
23 |
+
|
24 |
+
question="中国的首都是什么?"
|
25 |
+
input_text = "<s>Human: " + question + "\n</s><s>Assistant: "
|
26 |
+
input_ids = tokenizer(input_text, return_tensors='pt').input_ids.to(model.device)
|
27 |
+
|
28 |
+
with torch.no_grad():
|
29 |
+
with torch.autocast('cuda'):
|
30 |
+
output = model.generate(input_ids=input_ids,
|
31 |
+
max_new_tokens=max_new_tokens,
|
32 |
+
do_sample=True,
|
33 |
+
temperature=0.85,
|
34 |
+
top_k=None,
|
35 |
+
top_p=0.9,
|
36 |
+
use_cache=True,
|
37 |
+
**kwargs)
|
38 |
+
|
39 |
+
reply = tokenizer.decode(output[0], skip_special_tokens=False)
|
40 |
+
reply_return=reply.split('Assistant:')[-1].replace('</s>', '')
|
41 |
+
print('模型回答:', reply_return)
|
42 |
+
```
|
43 |
+
|
44 |
+
示例(多文档检索(问答)任务,输入文本大于10000字)
|
45 |
```
|
46 |
先阅读以下多个参考文档,然后根据文档内容,详细地回答问题,并指出所参考的文档的序号:
|
47 |
|
|
|
110 |
以上回答参考了文档-43。
|
111 |
```
|
112 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
113 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|