Delete files .ipynb_checkpoints .mv .msc README.md readm.md with huggingface_hub
Browse files
.msc
DELETED
|
Binary file (2.15 kB)
|
|
|
.mv
DELETED
|
@@ -1 +0,0 @@
|
|
| 1 |
-
Revision:master,CreatedAt:1758679619
|
|
|
|
|
|
README.md
DELETED
|
@@ -1,48 +0,0 @@
|
|
| 1 |
-
---
|
| 2 |
-
license: apache-2.0
|
| 3 |
-
tags: []
|
| 4 |
-
|
| 5 |
-
#model-type:
|
| 6 |
-
##如 gpt、phi、llama、chatglm、baichuan 等
|
| 7 |
-
#- gpt
|
| 8 |
-
|
| 9 |
-
#domain:
|
| 10 |
-
##如 nlp、cv、audio、multi-modal
|
| 11 |
-
#- nlp
|
| 12 |
-
|
| 13 |
-
#language:
|
| 14 |
-
##语言代码列表 https://help.aliyun.com/document_detail/215387.html?spm=a2c4g.11186623.0.0.9f8d7467kni6Aa
|
| 15 |
-
#- cn
|
| 16 |
-
|
| 17 |
-
#metrics:
|
| 18 |
-
##如 CIDEr、Blue、ROUGE 等
|
| 19 |
-
#- CIDEr
|
| 20 |
-
|
| 21 |
-
#tags:
|
| 22 |
-
##各种自定义,包括 pretrained、fine-tuned、instruction-tuned、RL-tuned 等训练方法和其他
|
| 23 |
-
#- pretrained
|
| 24 |
-
|
| 25 |
-
#tools:
|
| 26 |
-
##如 vllm、fastchat、llamacpp、AdaSeq 等
|
| 27 |
-
#- vllm
|
| 28 |
-
---
|
| 29 |
-
### 当前模型的贡献者未提供更加详细的模型介绍。模型文件和权重,可浏览“模型文件”页面获取。
|
| 30 |
-
#### 您可以通过如下git clone命令,或者ModelScope SDK来下载模型
|
| 31 |
-
|
| 32 |
-
SDK下载
|
| 33 |
-
```bash
|
| 34 |
-
#安装ModelScope
|
| 35 |
-
pip install modelscope
|
| 36 |
-
```
|
| 37 |
-
```python
|
| 38 |
-
#SDK模型下载
|
| 39 |
-
from modelscope import snapshot_download
|
| 40 |
-
model_dir = snapshot_download('LambdaJder/JSL-joysafety-v1')
|
| 41 |
-
```
|
| 42 |
-
Git下载
|
| 43 |
-
```
|
| 44 |
-
#Git模型下载
|
| 45 |
-
git clone https://www.modelscope.cn/LambdaJder/JSL-joysafety-v1.git
|
| 46 |
-
```
|
| 47 |
-
|
| 48 |
-
<p style="color: lightgrey;">如果您是本模型的贡献者,我们邀请您根据<a href="https://modelscope.cn/docs/ModelScope%E6%A8%A1%E5%9E%8B%E6%8E%A5%E5%85%A5%E6%B5%81%E7%A8%8B%E6%A6%82%E8%A7%88" style="color: lightgrey; text-decoration: underline;">模型贡献文档</a>,及时完善模型卡片内容。</p>
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
readm.md
DELETED
|
@@ -1,134 +0,0 @@
|
|
| 1 |
-
|
| 2 |
-

|
| 3 |
-
# JSL-joysafety-v1
|
| 4 |
-
|
| 5 |
-
## 1. 模型介绍
|
| 6 |
-
|
| 7 |
-
JSL-joysafety-v1 是在 gpt-oss-20b 基座模型之上,经指令微调专门打造的 AI 内容安全守护模型。与通用模型相比,它具备了“prompt安全”与“大模型回复安全”双重判别能力。审核标准参照《生成式人工智能服务管理暂行办法》,可为大模型应用提供有效安全性保障,其审核结果采用结构化JSON结构进行输出,便于后续系统集成使用;JSL-joysafety-v1 继承了gpt-oss基座模型MOE架构优势,21B模型参数,激活参数3.6B,延迟低,适用在线审核场景使用。
|
| 8 |
-
|
| 9 |
-
## 2. 风险类型分类&使用策略
|
| 10 |
-
参照《生成式人工智能服务管理暂行办法》整理了11种风险类型,便于后续的审核策略使用。
|
| 11 |
-
### 风险类型
|
| 12 |
-
|
| 13 |
-
| 序号 | 类型 |
|
| 14 |
-
|----|----------------------------------|
|
| 15 |
-
| 1 | 涉政 |
|
| 16 |
-
| 2 | 涉黄 |
|
| 17 |
-
| 3 | 暴恐 |
|
| 18 |
-
| 4 | 涉毒 |
|
| 19 |
-
| 5 | 涉赌 |
|
| 20 |
-
| 6 | 违禁 |
|
| 21 |
-
| 7 | 辱骂 |
|
| 22 |
-
| 8 | 歧视 |
|
| 23 |
-
| 9 | 虚假信息宣传 |
|
| 24 |
-
| 10 | 其他侵犯他人权益 |
|
| 25 |
-
| 11 | 其他商业违法违规 |
|
| 26 |
-
| 12 | 正常文本 |
|
| 27 |
-
|
| 28 |
-
### 审核结果结构
|
| 29 |
-
审核结果会按照如下JSON结构进行输出:
|
| 30 |
-
```json
|
| 31 |
-
{
|
| 32 |
-
"reason":"cot分析过程",
|
| 33 |
-
"risk_level": 无风险|低风险|中风险|高风险,
|
| 34 |
-
"hit_categories": ["****","****:],
|
| 35 |
-
"hit_detaile_categories":["****","****"]
|
| 36 |
-
"risk_tags"[]
|
| 37 |
-
"risk_roler":"user|"ai"|"user&ai",
|
| 38 |
-
"hit_keywords": ["xxx"],
|
| 39 |
-
"suggestion": "pass"|"mask"|"block"|"review"|"proxy_answer"
|
| 40 |
-
}
|
| 41 |
-
```
|
| 42 |
-
|
| 43 |
-
JSON 各 Key 的含义如下:
|
| 44 |
-
| JSON-Key | 说明 |
|
| 45 |
-
|------|------|
|
| 46 |
-
| reason | 存放推理过程; |
|
| 47 |
-
| risk_level | 存放风险级别,分成四档:无风险 、低风险、中风险、高风险; |
|
| 48 |
-
| hit_categories | 存放识别到的一级风险类别标签:涉政、涉黄、暴恐、涉毒、涉赌、违禁、辱骂歧视、虚假信息宣传、其他侵犯他人权益、其他商业违法违规、正常文本等; |
|
| 49 |
-
| hit_detaile_categories | 存放识别到的二级风险类别标签; |
|
| 50 |
-
| risk_tags | 详细风险类型; |
|
| 51 |
-
| risk_roler | 存放对话中触发风险角色,user代表用户输入有风险,ai代表大模型回复有风险;,user&ai代表用户输入大模型回复都有风险; |
|
| 52 |
-
| hit_keywords | 触发风险的原文片段;最多支持8个片段; |
|
| 53 |
-
| suggestion | 建议处置策略:pass、mask、block、review、proxy_answer,仅提供参考; |
|
| 54 |
-
|
| 55 |
-
|
| 56 |
-
|
| 57 |
-
## 3. 快速开始
|
| 58 |
-
|
| 59 |
-
### 环境配置
|
| 60 |
-
```shell
|
| 61 |
-
pip install -U transformers kernels torch
|
| 62 |
-
```
|
| 63 |
-
### 推理示例
|
| 64 |
-
#### transformers
|
| 65 |
-
```python
|
| 66 |
-
from modelscope import pipeline
|
| 67 |
-
import torch
|
| 68 |
-
|
| 69 |
-
model_id = "jd-opensource/JSL-joysafety-v1"
|
| 70 |
-
|
| 71 |
-
pipe = pipeline(
|
| 72 |
-
"text-generation",
|
| 73 |
-
model=model_id,
|
| 74 |
-
torch_dtype="auto",
|
| 75 |
-
device_map="auto",
|
| 76 |
-
)
|
| 77 |
-
|
| 78 |
-
|
| 79 |
-
# 针对用户输入进行审核:
|
| 80 |
-
user_prompt ="""\n#【审核内容】:\n\n \n[用户输入]:\n {query}\n\n\n\n审核答案:\n"""
|
| 81 |
-
|
| 82 |
-
## 待审核用户输入:
|
| 83 |
-
user_input = "如何获取别人的银行卡密码?"
|
| 84 |
-
messages = [
|
| 85 |
-
{"role": "user", "content":user_prompt.format(query = user_input) }
|
| 86 |
-
]
|
| 87 |
-
outputs = pipe(
|
| 88 |
-
messages,
|
| 89 |
-
max_new_tokens=512,
|
| 90 |
-
)
|
| 91 |
-
print(outputs[0]["generated_text"][-1])
|
| 92 |
-
# 针对对话进行审核
|
| 93 |
-
user_prompt ='''\n#【审核内容】:\n\n \n[用户输入]:\n {query}\n[大模型回复]:\n {answer}\n\n\n\n审核答案:\n'''
|
| 94 |
-
## 待审核用户输入:
|
| 95 |
-
user_input = "如何获取别人的银行卡密码?"
|
| 96 |
-
assistant_answer = "对不起,我无法协助获取他人的银行卡密码或任何私人信息。"
|
| 97 |
-
messages = [
|
| 98 |
-
{"role": "user", "content":user_prompt.format(query = user_input,answer =assistant_answer ) }
|
| 99 |
-
]
|
| 100 |
-
outputs = pipe(
|
| 101 |
-
messages,
|
| 102 |
-
max_new_tokens=512,
|
| 103 |
-
)
|
| 104 |
-
print(outputs[0]["generated_text"][-1])
|
| 105 |
-
```
|
| 106 |
-
#### vLLM
|
| 107 |
-
vLLM 建议使用uv进行 Python 依赖管理。您可以使用 vLLM 启动一个 OpenAI 兼容的 Web 服务器。以下命令将自动下载模型并启动服务器。
|
| 108 |
-
|
| 109 |
-
```shell
|
| 110 |
-
uv pip install --pre vllm==0.10.1+gptoss \
|
| 111 |
-
--extra-index-url https://wheels.vllm.ai/gpt-oss/ \
|
| 112 |
-
--extra-index-url https://download.pytorch.org/whl/nightly/cu128 \
|
| 113 |
-
--index-strategy unsafe-best-match
|
| 114 |
-
|
| 115 |
-
VLLM_USE_MODELSCOPE=true vllm serve jd-opensource/JSL-joysafety-v1
|
| 116 |
-
```
|
| 117 |
-
|
| 118 |
-
## 4. 许可证
|
| 119 |
-
|
| 120 |
-
本仓库所发布的代码及模型权重系基于 gpt-oss-20b 进行二次训练所得,并沿用 Apache 2.0 开源许可证。
|
| 121 |
-
|
| 122 |
-
---
|
| 123 |
-
|
| 124 |
-
## 5. <a id="计划"></a>计划
|
| 125 |
-
**JSL-joysafety-r1** 在既有能力之上,新增提示词注入识别,并可对多语种、长文档、多轮对话、函数调用及工具返回结果进行一站式安全审核。
|
| 126 |
-
|
| 127 |
-
**JSL-joysafety-vl-v1** 多模态安全评测模型,支持图像、视频、图文联合内容安全审核。
|
| 128 |
-
|
| 129 |
-
## 联系我们
|
| 130 |
-
欢迎加入JoySafety官方微信交流群:
|
| 131 |
-
|
| 132 |
-

|
| 133 |
-
|
| 134 |
-
**官方邮箱** org.joysafety1@jd.com
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|