Skip to content

修改语言检测文档,涉政模型文档,敏感词代码及文档 #426

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
darkrush merged 2 commits into from
Jun 19, 2025
Merged

修改语言检测文档,涉政模型文档,敏感词代码及文档 #426

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
42b1e72
修改语言检测文档,涉政模型文档,敏感词代码及文档 (#19)
2471023025 May 22, 2025
a617447
bug fix (#16)
Adela-Yu-Coder May 22, 2025
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
88 changes: 87 additions & 1 deletion docs/llm_web_kit/model/lang_id.md
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -4,7 +4,7 @@

is_218e为True时使用lid218e模型，在多个小语种中有更好的表现，除个别容易使模型混淆的情况外，会返回正常的language_details字段，若该参数为False，则language_details字段为空，默认值为True

is_cn_specific为True时，会对文本中的中文文本进行细分，分为zho-Hans(简体中文)或zho-Hant(繁体中文)，结果在language_details字段中，默认值为False
is_cn_specific为True时，会对文本中的中文文本进行细分，分为zho-Hans(简体中文)或zho-Hant(繁体中文),结果在language_details字段中，默认值为False,如果需要使用，请先pip install langdetect_zh==1.0.4,该package使用langdetect的方法,并针对中文进行了特调,能有效识别简体中文和繁体中文

## 配置文件需要改动的部分

Expand Down Expand Up @@ -120,3 +120,89 @@ print(update_language_by_str(text, is_cn_specific=True))
总时间: 1.3538 秒

处理速度: 443.91 条/秒

## 性能说明

测试集使用gsarti/flores_101，该数据集包含102种语言的并行句子，每个语种2009条测试集路径：https://huggingface.co/datasets/gsarti/flores_101

下表所示lid176为单模型结果，模型路径为s3://web-parse-huawei/shared_resource/language/lid176.bin

lid218e也为单模型结果，模型路径为s3://web-parse-huawei/shared_resource/language/lid218e.bin

级联方案即为该代码调用方案，使用lid176判断zh, en, ja, ko，使用lid218e判断其他语种，使用langdetect_zh区分简体中文与繁体中文

该表统计了三种模型在102种语言上错误的次数，其中lid176繁体中文全错是考虑到该模型无法区分简体中文和繁体中文

| 级联方案 | | lid176 | | lid218e | |
| --------- | -------- | -------- | -------- | --------- | -------- |
| 真实语言 | 错误次数 | 真实语言 | 错误次数 | 真实语言 | 错误次数 |
| bos | 1079 | zho_trad | 2009 | bos | 1079 |
| kam | 767 | ful | 2009 | kam | 765 |
| zho_trad | 18 | lug | 2009 | zho_trad | 623 |
| hrv | 197 | hau | 2009 | zho_simpl | 229 |
| nya | 165 | ibo | 2009 | hrv | 197 |
| kea | 145 | kea | 2009 | nya | 161 |
| msa | 145 | kam | 2009 | kea | 145 |
| ful | 67 | lin | 2009 | msa | 145 |
| xho | 51 | luo | 2009 | ful | 56 |
| umb | 46 | mri | 2009 | umb | 46 |
| zul | 38 | nso | 2009 | jpn | 46 |
| fas | 38 | nya | 2009 | fas | 38 |
| ind | 37 | orm | 2009 | ind | 37 |
| mri | 27 | sna | 2009 | xho | 32 |
| wol | 22 | umb | 2009 | zul | 16 |
| ast | 16 | wol | 2009 | ast | 16 |
| dan | 13 | xho | 2009 | dan | 13 |
| nob | 13 | zul | 2009 | wol | 13 |
| nso | 12 | bos | 1879 | nob | 13 |
| luo | 11 | ast | 1373 | nso | 11 |
| lug | 11 | som | 1184 | luo | 8 |
| jav | 9 | msa | 1131 | lug | 7 |
| sna | 9 | yor | 943 | pus | 7 |
| ibo | 8 | oci | 753 | glg | 6 |
| afr | 7 | hrv | 609 | jav | 6 |
| pus | 7 | jav | 590 | mri | 5 |
| glg | 6 | afr | 574 | hin | 4 |
| som | 4 | glg | 294 | swe | 3 |
| swh | 4 | uzb | 188 | yor | 3 |
| hin | 4 | ltz | 151 | lin | 3 |
| yor | 4 | ceb | 144 | lao | 2 |
| lin | 4 | nob | 137 | oci | 2 |
| ceb | 3 | swh | 118 | som | 2 |
| swe | 3 | mlt | 109 | ceb | 2 |
| lao | 2 | dan | 90 | khm | 2 |
| oci | 2 | slv | 56 | slv | 1 |
| uzb | 2 | ind | 48 | uzb | 1 |
| orm | 2 | slk | 41 | npi | 1 |
| nld | 2 | pus | 37 | tgl | 1 |
| hau | 2 | gle | 26 | bul | 1 |
| slv | 1 | npi | 18 | fra | 1 |
| zho_simpl | 1 | azj | 17 | hau | 1 |
| npi | 1 | asm | 14 | ita | 1 |
| eng | 1 | tgk | 13 | ltz | 1 |
| tgl | 1 | isl | 13 | kaz | 1 |
| est | 1 | est | 12 | por | 1 |
| bul | 1 | snd | 12 | afr | 1 |
| fra | 1 | cym | 11 | spa | 1 |
| ita | 1 | cat | 11 | | |
| khm | 1 | srp | 11 | | |
| ltz | 1 | kir | 10 | | |
| kaz | 1 | nld | 5 | | |
| por | 1 | por | 5 | | |
| spa | 1 | swe | 4 | | |
| | | mkd | 4 | | |
| | | lav | 4 | | |
| | | urd | 3 | | |
| | | tgl | 3 | | |
| | | kaz | 2 | | |
| | | ron | 2 | | |
| | | ita | 2 | | |
| | | bel | 2 | | |
| | | bul | 2 | | |
| | | lit | 2 | | |
| | | lao | 1 | | |
| | | ckb | 1 | | |

根据统计表格，lid176准确率0.7715，lid218e准确率为0.9817，级联方案准确率为0.9853，准确率公式为：1-sum(错误次数)/(102\*2009)

级联方案相比于lid176提升了多语种的准确率，同时也解决了lid218e针对部分语种（中文简体、中文繁体、日语）的错误
46 changes: 36 additions & 10 deletions docs/llm_web_kit/model/politics_detector.md
View file
Open in desktop
Original file line number Diff line number Diff line change
@@ -1,8 +1,8 @@
## 作用

识别中文或英文文本中的涉政内容，目前包含了新旧两类接口，旧的接口接收单条数据，并返回该数据的涉政分数，分数接近1代表不涉政，分数接近0则代表涉政。目前旧的接口仅支持CPU模型。
识别中文或英文文本中的涉政内容，目前包含了新旧两类接口，25m3_cpu模型接口接收单条数据，并返回该数据的涉政分数，分数接近1代表不涉政，分数接近0则代表涉政。目前25m3_cpu模型接口仅支持CPU模型。

新的接口检测结果以ModelResponse类返回，该类包含is_remained和details两个字段，其中is_remained代表数据是否需要保留，details则是一个包含涉政分数等详细信息的字典。新的接口支持CPU和GPU两种模型。
25m3模型接口检测结果以ModelResponse类返回，该类包含is_remained和details两个字段，其中is_remained代表数据是否需要保留，details则是一个包含涉政分数等详细信息的字典。25m3模型接口支持GPU模型。

## 配置文件需要改动的部分

Expand All @@ -13,20 +13,20 @@
"common":{
"cache_path": "~/.llm_web_kit_cache"
},
"political-24m7":{
"download_path": "s3://web-parse-huawei/shared_resource/political/24m7.zip",
"md5": "97eabb56268a3af3f68e8a96a50d5f80",
},
"political-25m3":{
"download_path": "s3://web-parse-huawei/shared_resource/political/25m3.zip",
"md5": "d0d14a561f987763d654165b536b5858",
},
"political-25m3_cpu":{
"download_path": "s3://web-parse-huawei/shared_resource/political/25m3_cpu.zip",
"md5": "926359a393de6a36c1b4be403711767f",
},
},
```

## 调用方法

1. 旧的接口调用方法如下：
1. 25m3_cpu模型接口调用方法如下：

```python
from llm_web_kit.model.politics_detector import *
Expand Down Expand Up @@ -81,7 +81,7 @@ print(political_filter_cpu(text, "en"))
# 输出结果为：{'political_prob': 1.0000100135803223}
```

2. 新的接口调用方法如下：
2. 25m3模型接口调用方法如下：

```python
from llm_web_kit.model.model_impl import ModelFactory, ModelType, DeviceType
Expand Down Expand Up @@ -113,7 +113,7 @@ for i in range(0, len(requests), batch_size):

## 运行时间

1. 旧的接口（political_filter_cpu）
1. 25m3_cpu模型接口（political_filter_cpu）

使用型号为`AMD EPYC 7742`的cpu单核进行测试，测试集总共有 77861 条数据（均是中英文的数据），下面只统计了political_filter_cpu接口本身的耗时，排除了数据读取的时间。

Expand All @@ -127,7 +127,7 @@ for i in range(0, len(requests), batch_size):

每秒可处理: 416.3049条数据

2. 新的接口（predictor.predict_batch）
2. 25m3模型接口（predictor.predict_batch）

使用单卡NVIDIA A100测试涉政的GPU模型，测试集共有39111条数据，下面统计了不同batch_size下，predictor.predict_batch接口的速度，该接口内部包括tokenize和模型推理操作。

Expand Down Expand Up @@ -159,3 +159,29 @@ for i in range(0, len(requests), batch_size):
| 128 | 31.580092769179686 |
| 256 | 24.26296225431703 |
| 512 | cuda out of memory |

## 性能说明

25m3_cpu模型（threshold=0.5）：

测试集路径：s3://xyz-process-ylk2/xyz-users/huyucheng1/political_data_202502/test/

| 指标 | 新模型值 | 旧模型值 |
| ------------- | -------------------- | -------------------- |
| **F1** | 0.9089603520041284 | 0.8831507760632497 |
| **Accuracy** | 0.8624864742896118 | 0.8013861609546715 |
| **Precision** | 0.9041776426882809 | 0.7913184992146802 |
| **Recall** | 0.9137939273134369 | 0.999095513748191 |
| **TN** | 68641 | 19820 |
| **FP** | 28373 | 77194 |
| **FN** | 25257 | 265 |
| **TP** | 267727 | 292719 |
| **Prec_Pos** | 0.9041776426882809 | 0.7913184992146802 |
| **Recl_Pos** | 0.9137939273134369 | 0.999095513748191 |
| **F1_Pos** | 0.9089603520041284 | 0.8831507760632497 |
| **Prec_Neg** | 0.7310166350720995 | 0.986806074184715 |
| **Recl_Neg** | 0.7075370565073082 | 0.204300410250067 |
| **F1_Neg** | 0.719085232986926 | 0.3385169813576546 |
| **qps** | 1493.477337807 条/秒 | 1674.157845704 条/秒 |

注：上述指标均是在集群中得出，单核运行时间请参考运行时间第一小节
30 changes: 28 additions & 2 deletions docs/llm_web_kit/model/rule_based_safety_module.md
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -10,8 +10,8 @@
"cache_path": "~/.llm_web_kit_cache"
},
"unsafe_words":{
"download_path": "s3://web-parse-huawei/shared_resource/political/unsafe_words.jsonl",
"md5": "e81dd1050a79f68b9d9b3f66baadde66",
"download_path": "s3://web-parse-huawei/shared_resource/unsafe_words/unsafe_words_porn_politics.jsonl",
"md5": "ef51faf114353d987ec97b211a8d2b06",
},
"xyz_internal_unsafe_words":{
"download_path": "s3://web-parse-huawei/shared_resource/political/xyz_internal_unsafe_words.jsonl",
Expand Down Expand Up @@ -51,6 +51,32 @@ m.process("your content",
'safety_infos': {'domain_level': '', 'hit_unsafe_words': False}}
```

### 敏感词检测模块用法示例

```python
from llm_web_kit.model.unsafe_words_detector import *

checker = UnsafeWordChecker(language="zh-en")

content = "64式销售QQ"
unsafe_words = checker.check_unsafe_words(
content_str=content,
)
print(unsafe_words)
[{'word': '64式', 'type': '违禁品', 'level': 'L3', 'language': 'zh', 'count': 1.0}, {'word': '64式销售', 'type': '违禁品', 'level': 'L3', 'language': 'zh', 'count': 1.0}, {'word': '销售', 'type': '广告营销', 'level': 'L3', 'language': 'zh', 'count': 1.0}, {'word': '64式销售qq', 'type': '违禁品', 'level': 'L1', 'language': 'zh', 'count': 1.0}]

checker = UnsafeWordsFilter()
content = "64式销售QQ"
#from_safe_source:是否来自安全来源。如果是，直接返回安全。
#from_domestic_source: 是否来自国内来源。如果是，仅检查 L1 级别的不安全词；否则检查 L1 和 L2 级别。
result = checker.filter(
content,
'zh',
from_safe_source = False,
from_domestic_source = True,
)
```

## 速度

### 整体速度：
Expand Down
2 changes: 1 addition & 1 deletion llm_web_kit/model/model_impl.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -112,7 +112,7 @@ def convert_result_to_response(self, result: dict) -> ModelResponse:
# raise NotImplementedError
# TODO convert result to response ensure the threshold
return PoliticalResponse(
is_remained=result['political_prob'] > 0.99, details=result
is_remained=result['political_prob'] > 0.89, details=result
)


Expand Down
20 changes: 9 additions & 11 deletions llm_web_kit/model/politics_detector.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -27,20 +27,20 @@ def __init__(self, model_path: str = None):
if not model_path:
model_path = self.auto_download()
model_bin_path = os.path.join(model_path, 'model.bin')
tokenizer_path = os.path.join(model_path, 'internlm2-chat-20b')
tokenizer_path = os.path.join(model_path, 'qwen2.5_7b_tokenizer')

self.model = fasttext.load_model(model_bin_path)
self.tokenizer = transformer.AutoTokenizer.from_pretrained(
tokenizer_path, use_fast=False, trust_remote_code=True
)

def auto_download(self):
"""Default download the 24m7.zip model."""
resource_name = 'political-24m7'
"""Default download the 25m3_cpu.zip model."""
resource_name = 'political-25m3_cpu'
resource_config = load_config()['resources']
political_24m7_config: dict = resource_config[resource_name]
political_24m7_s3 = political_24m7_config['download_path']
political_24m7_md5 = political_24m7_config.get('md5', '')
political_25m3_cpu_config: dict = resource_config[resource_name]
political_25m3_cpu_s3 = political_25m3_cpu_config['download_path']
political_25m3_cpu_md5 = political_25m3_cpu_config.get('md5', '')
# get the zip path calculated by the s3 path
zip_path = os.path.join(CACHE_DIR, f'{resource_name}.zip')
# the unzip path is calculated by the zip path
Expand All @@ -52,9 +52,9 @@ def auto_download(self):
logger.info(f'try to unzip from zip_path: {zip_path}')
if not os.path.exists(zip_path):
logger.info(f'zip_path: {zip_path} does not exist')
logger.info(f'downloading {political_24m7_s3}')
logger.info(f'downloading {political_25m3_cpu_s3}')
zip_path = download_auto_file(
political_24m7_s3, zip_path, political_24m7_md5
political_25m3_cpu_s3, zip_path, political_25m3_cpu_md5
)
logger.info(f'unzipping {zip_path}')
unzip_path = unzip_local_file(zip_path, unzip_path)
Expand Down Expand Up @@ -195,7 +195,7 @@ def get_singleton_political_detect() -> PoliticalDetector:
def decide_political_by_prob(
predictions: Tuple[str], probabilities: Tuple[float]
) -> float:
idx = predictions.index('__label__normal')
idx = predictions.index('__label__positive')
normal_score = probabilities[idx]
return float(normal_score)

Expand Down Expand Up @@ -226,8 +226,6 @@ def political_filter_cpu(data_dict: Dict[str, Any], language: str):

if __name__ == '__main__':
test_cases = []
test_cases.append('你好，我很高兴见到你！')
test_cases.append('hello, nice to meet you!')
test_cases.append('你好，唔該幫我一個忙？')
test_cases.append('Bawo ni? Mo nife Yoruba. ')
test_cases.append(
Expand Down
2 changes: 1 addition & 1 deletion llm_web_kit/model/porn_detector.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -172,7 +172,7 @@ def __init__(self, model_path: str = None) -> None:
model_config = json.load(reader)
self.cls_index = int(model_config.get('cls_index', 1))
self.use_sigmoid = bool(model_config.get('use_sigmoid', False))
self.max_tokens = int(model_config.get('max_tokens', 300))
self.max_tokens = int(model_config.get('max_tokens', 512))
self.remain_tail = min(
self.max_tokens - 1, int(model_config.get('remain_tail', -1))
)
Expand Down
6 changes: 4 additions & 2 deletions llm_web_kit/model/unsafe_words_detector.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -68,7 +68,7 @@
unsafe_words_file_path = auto_download(language)
t2 = time.time()
print(
f'-----------------auto_download cost time: {t2-t1} , language: {language}------------------'
f'-----------------auto_download cost time: {t2 - t1} , language: {language}------------------'
)
with open(unsafe_words_file_path, 'r') as f:
lines = f.readlines()
Expand All @@ -85,6 +85,8 @@
words = {}
for line in lines:
w = json_loads(line)
if w.get('tag') == 'delete':
continue

Check warning on line 89 in llm_web_kit/model/unsafe_words_detector.py

View check run for this annotation

Codecov / codecov/patch

llm_web_kit/model/unsafe_words_detector.py#L89 

Added line #L89 was not covered by tests
word = str(w.get('word') or '').lower()
if not word:
continue
Expand Down Expand Up @@ -163,7 +165,7 @@
self.ac = get_ac(language)
t2 = time.time()
print(
f'---------------UnsafeWordChecker init time: {t2-t1} , language: {language}-----------------'
f'---------------UnsafeWordChecker init time: {t2 - t1} , language: {language}-----------------'
)

def check_unsafe_words(self, content_str: str) -> list:
Expand Down
8 changes: 4 additions & 4 deletions tests/llm_web_kit/model/test_model_impl.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -92,14 +92,14 @@ def test_convert_result_to_response(self, mock_load_model):
mock_load_model.return_value = MagicMock()
model = PoliticalCPUModel()

# Test case where political_prob > 0.99 (should be flagged)
result = {'political_prob': 0.995}
# Test case where political_prob > 0.89 (should be flagged)
result = {'political_prob': 0.9}
response = model.convert_result_to_response(result)
assert response.is_remained
assert response.details == result

# Test case where political_prob <= 0.99 (should not be flagged)
result = {'political_prob': 0.985}
# Test case where political_prob <= 0.89 (should not be flagged)
result = {'political_prob': 0.88}
response = model.convert_result_to_response(result)
assert not response.is_remained
assert response.details == result
Expand Down
Loading