Create README.md
Browse files
README.md
CHANGED
@@ -1,3 +1,59 @@
|
|
1 |
---
|
2 |
license: apache-2.0
|
|
|
|
|
3 |
---
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
---
|
2 |
license: apache-2.0
|
3 |
+
datasets:
|
4 |
+
- w8ay/security-paper-datasets
|
5 |
---
|
6 |
+
|
7 |
+
## 使用
|
8 |
+
商业模型对于网络安全领域问题大多会有道德限制,所以基于网络安全数据训练了一个模型,模型基于Baichuan 13B,模型参数大小130亿,至少需要30G显存运行,35G最佳。
|
9 |
+
- transformers
|
10 |
+
- peft
|
11 |
+
|
12 |
+
**模型加载**
|
13 |
+
```python
|
14 |
+
from transformers import AutoModelForCausalLM, AutoTokenizer
|
15 |
+
import torch
|
16 |
+
from peft import PeftModel
|
17 |
+
device = 'auto'
|
18 |
+
tokenizer = AutoTokenizer.from_pretrained("w8ay/secgpt", trust_remote_code=True)
|
19 |
+
model = AutoModelForCausalLM.from_pretrained("w8ay/secgpt",
|
20 |
+
trust_remote_code=True,
|
21 |
+
device_map=device,
|
22 |
+
torch_dtype=torch.float16)
|
23 |
+
print("模型加载成功")
|
24 |
+
|
25 |
+
```
|
26 |
+
**调用**
|
27 |
+
```python
|
28 |
+
def reformat_sft(instruction, input):
|
29 |
+
if input:
|
30 |
+
prefix = (
|
31 |
+
"Below is an instruction that describes a task, paired with an input that provides further context. "
|
32 |
+
"Write a response that appropriately completes the request.\n"
|
33 |
+
f"### Instruction:\n{instruction}\n\n### Input:\n{input}\n\n### Response:"
|
34 |
+
)
|
35 |
+
else:
|
36 |
+
prefix = (
|
37 |
+
"Below is an instruction that describes a task. "
|
38 |
+
"Write a response that appropriately completes the request.\n"
|
39 |
+
f"### Instruction:\n{instruction}\n\n### Response:"
|
40 |
+
)
|
41 |
+
return prefix
|
42 |
+
|
43 |
+
query = '''介绍sqlmap如何使用'''
|
44 |
+
query = reformat_sft(query,'')
|
45 |
+
|
46 |
+
generation_kwargs = {
|
47 |
+
"top_p": 0.7,
|
48 |
+
"temperature": 0.3,
|
49 |
+
"max_new_tokens": 2000,
|
50 |
+
"do_sample": True,
|
51 |
+
"repetition_penalty":1.1
|
52 |
+
}
|
53 |
+
|
54 |
+
inputs = tokenizer.encode(query, return_tensors='pt', truncation=True)
|
55 |
+
inputs = inputs.cuda()
|
56 |
+
generate = model.generate(input_ids=inputs, **generation_kwargs)
|
57 |
+
output = tokenizer.decode(generate[0])
|
58 |
+
print(output)
|
59 |
+
```
|