Spaces:
Runtime error
Runtime error
gabrielchua
commited on
Commit
•
176cc29
1
Parent(s):
a573698
Update to lionguard-v1
Browse files
app.py
CHANGED
@@ -1,82 +1,48 @@
|
|
1 |
-
"""
|
2 |
-
app.py
|
3 |
-
"""
|
4 |
-
|
5 |
import json
|
6 |
-
from typing import
|
7 |
|
8 |
import gradio as gr
|
9 |
import numpy as np
|
10 |
import pandas as pd
|
11 |
-
import spaces
|
12 |
import torch
|
13 |
from huggingface_hub import hf_hub_download
|
14 |
-
from sklearn.linear_model import RidgeClassifier
|
15 |
from transformers import AutoModel, AutoTokenizer
|
|
|
16 |
|
17 |
-
# Define the
|
18 |
-
|
19 |
-
"dsaidgovsg/Lionguard-binary-v1.0",
|
20 |
-
"dsaidgovsg/Lionguard-harassment-v1.0",
|
21 |
-
"dsaidgovsg/Lionguard-hateful-v1.0",
|
22 |
-
"dsaidgovsg/Lionguard-public_harm-v1.0",
|
23 |
-
"dsaidgovsg/Lionguard-self_harm-v1.0",
|
24 |
-
"dsaidgovsg/Lionguard-sexual-v1.0",
|
25 |
-
"dsaidgovsg/Lionguard-toxic-v1.0",
|
26 |
-
"dsaidgovsg/Lionguard-violent-v1.0",
|
27 |
-
]
|
28 |
|
29 |
-
def load_config(
|
30 |
"""
|
31 |
-
Load the configuration for
|
32 |
-
|
33 |
-
Args:
|
34 |
-
model_repo (str): The model repository name.
|
35 |
-
|
36 |
Returns:
|
37 |
Dict[str, Any]: The configuration dictionary.
|
38 |
"""
|
39 |
-
config_path = hf_hub_download(repo_id=
|
40 |
with open(config_path, 'r') as f:
|
41 |
return json.load(f)
|
42 |
|
43 |
-
def load_all_configs() -> Dict[str, Dict[str, Any]]:
|
44 |
-
"""
|
45 |
-
Load configurations for all Lionguard models.
|
46 |
-
|
47 |
-
Returns:
|
48 |
-
Dict[str, Dict[str, Any]]: A dictionary of model configurations.
|
49 |
-
"""
|
50 |
-
model_configs = {}
|
51 |
-
for model_repo in Lionguard_models:
|
52 |
-
model_configs[model_repo] = load_config(model_repo)
|
53 |
-
print("All model configurations loaded.")
|
54 |
-
return model_configs
|
55 |
-
|
56 |
-
@spaces.GPU
|
57 |
def get_embeddings(device: str, data: List[str], config: Dict[str, Any]) -> np.ndarray:
|
58 |
"""
|
59 |
Generate embeddings for the input data using the specified model configuration.
|
60 |
-
|
61 |
Args:
|
62 |
device (str): The device to use for computations.
|
63 |
data (List[str]): The input text data.
|
64 |
config (Dict[str, Any]): The model configuration.
|
65 |
-
|
66 |
Returns:
|
67 |
np.ndarray: The generated embeddings.
|
68 |
"""
|
69 |
-
tokenizer = AutoTokenizer.from_pretrained(config['tokenizer'])
|
70 |
-
model = AutoModel.from_pretrained(config['
|
71 |
model.eval()
|
72 |
model.to(device)
|
73 |
|
74 |
-
batch_size = config['batch_size']
|
75 |
num_batches = int(np.ceil(len(data)/batch_size))
|
76 |
output = []
|
77 |
for i in range(num_batches):
|
78 |
sentences = data[i*batch_size:(i+1)*batch_size]
|
79 |
-
encoded_input = tokenizer(sentences, max_length=config['max_length'], padding=True, truncation=True, return_tensors='pt')
|
80 |
encoded_input.to(device)
|
81 |
with torch.no_grad():
|
82 |
model_output = model(**encoded_input)
|
@@ -85,107 +51,85 @@ def get_embeddings(device: str, data: List[str], config: Dict[str, Any]) -> np.n
|
|
85 |
output.extend(sentence_embeddings.cpu().numpy())
|
86 |
return np.array(output)
|
87 |
|
88 |
-
def
|
89 |
"""
|
90 |
-
|
91 |
-
|
92 |
-
Args:
|
93 |
-
model (RidgeClassifier): The Ridge Classifier model.
|
94 |
-
attributes (Dict[str, Any]): The attributes to set.
|
95 |
-
|
96 |
-
Returns:
|
97 |
-
RidgeClassifier: The updated Ridge Classifier model.
|
98 |
-
"""
|
99 |
-
model.coef_ = np.array(attributes['coef_'])
|
100 |
-
model.intercept_ = np.array(attributes['intercept_'])
|
101 |
-
model.n_features_in_ = np.array(attributes['n_features_in_'])
|
102 |
-
return model
|
103 |
-
|
104 |
-
def convert_decision_to_proba(d: np.ndarray) -> np.ndarray:
|
105 |
-
"""
|
106 |
-
Convert decision function values to probabilities.
|
107 |
-
|
108 |
Args:
|
109 |
-
|
110 |
-
|
111 |
Returns:
|
112 |
-
|
113 |
"""
|
114 |
-
|
115 |
-
|
116 |
-
|
117 |
-
|
118 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
119 |
"""
|
120 |
-
Predict
|
121 |
-
|
122 |
Args:
|
123 |
text (str): The input text to predict on.
|
124 |
-
|
125 |
-
|
126 |
Returns:
|
127 |
-
pd.DataFrame: A DataFrame containing prediction
|
128 |
"""
|
129 |
if not text.strip():
|
130 |
return None
|
131 |
|
132 |
-
|
133 |
-
|
134 |
-
|
135 |
-
|
136 |
-
|
137 |
-
|
138 |
-
|
139 |
-
|
140 |
-
|
141 |
-
|
142 |
-
print(f"Configuration for {model_repo} not found. Skipping...")
|
143 |
-
continue
|
144 |
-
|
145 |
-
config = model_configs[model_repo]
|
146 |
-
|
147 |
-
model_fp = hf_hub_download(repo_id=model_repo, filename=config['model_name'])
|
148 |
-
with open(model_fp, 'r') as json_file:
|
149 |
-
model_params = json.load(json_file)
|
150 |
|
151 |
-
|
152 |
-
model_attributes = model_params["attributes"]
|
153 |
-
model_params.pop("attributes", None)
|
154 |
-
model.set_params(**model_params)
|
155 |
-
model = set_model_atttributes(model, model_attributes)
|
156 |
-
|
157 |
-
preds = convert_decision_to_proba(model.decision_function(embeddings_df))[:,1]
|
158 |
-
|
159 |
-
model_name = model_repo.split('/')[-1].split('-')[1]
|
160 |
-
results.append({"Category": model_name, "Probability": float(preds[0])})
|
161 |
|
162 |
-
|
163 |
-
|
164 |
-
if result_df.shape[0] > 0:
|
165 |
-
return result_df
|
166 |
-
else:
|
167 |
-
return None
|
168 |
-
|
169 |
-
def create_interface(model_configs: Dict[str, Dict[str, Any]]) -> gr.Interface:
|
170 |
"""
|
171 |
Create the Gradio interface for the Lionguard demo.
|
172 |
-
|
173 |
Args:
|
174 |
-
|
175 |
-
|
176 |
Returns:
|
177 |
gr.Interface: The Gradio interface object.
|
178 |
"""
|
179 |
return gr.Interface(
|
180 |
-
fn=lambda text:
|
181 |
inputs=gr.Textbox(lines=3, placeholder="Enter text here..."),
|
182 |
-
outputs=gr.DataFrame(label="Prediction
|
183 |
title="🦁 Lionguard Demo",
|
184 |
description="Lionguard is a Singapore-contextualized moderation classifier that can serve against unsafe LLM outputs.",
|
185 |
allow_flagging="never"
|
186 |
)
|
187 |
|
188 |
if __name__ == "__main__":
|
189 |
-
|
190 |
-
iface = create_interface(
|
191 |
iface.launch()
|
|
|
|
|
|
|
|
|
|
|
1 |
import json
|
2 |
+
from typing import List, Dict, Any
|
3 |
|
4 |
import gradio as gr
|
5 |
import numpy as np
|
6 |
import pandas as pd
|
|
|
7 |
import torch
|
8 |
from huggingface_hub import hf_hub_download
|
|
|
9 |
from transformers import AutoModel, AutoTokenizer
|
10 |
+
import onnxruntime as rt
|
11 |
|
12 |
+
# Define the Lionguard model repository
|
13 |
+
REPO_PATH = "govtech/lionguard-v1"
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
14 |
|
15 |
+
def load_config() -> Dict[str, Any]:
|
16 |
"""
|
17 |
+
Load the configuration for the Lionguard model.
|
|
|
|
|
|
|
|
|
18 |
Returns:
|
19 |
Dict[str, Any]: The configuration dictionary.
|
20 |
"""
|
21 |
+
config_path = hf_hub_download(repo_id=REPO_PATH, filename="config.json")
|
22 |
with open(config_path, 'r') as f:
|
23 |
return json.load(f)
|
24 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
25 |
def get_embeddings(device: str, data: List[str], config: Dict[str, Any]) -> np.ndarray:
|
26 |
"""
|
27 |
Generate embeddings for the input data using the specified model configuration.
|
|
|
28 |
Args:
|
29 |
device (str): The device to use for computations.
|
30 |
data (List[str]): The input text data.
|
31 |
config (Dict[str, Any]): The model configuration.
|
|
|
32 |
Returns:
|
33 |
np.ndarray: The generated embeddings.
|
34 |
"""
|
35 |
+
tokenizer = AutoTokenizer.from_pretrained(config['embedding']['tokenizer'])
|
36 |
+
model = AutoModel.from_pretrained(config['embedding']['model'])
|
37 |
model.eval()
|
38 |
model.to(device)
|
39 |
|
40 |
+
batch_size = config['embedding']['batch_size']
|
41 |
num_batches = int(np.ceil(len(data)/batch_size))
|
42 |
output = []
|
43 |
for i in range(num_batches):
|
44 |
sentences = data[i*batch_size:(i+1)*batch_size]
|
45 |
+
encoded_input = tokenizer(sentences, max_length=config['embedding']['max_length'], padding=True, truncation=True, return_tensors='pt')
|
46 |
encoded_input.to(device)
|
47 |
with torch.no_grad():
|
48 |
model_output = model(**encoded_input)
|
|
|
51 |
output.extend(sentence_embeddings.cpu().numpy())
|
52 |
return np.array(output)
|
53 |
|
54 |
+
def predict(text: str, config: Dict[str, Any]) -> Dict[str, Dict[str, Any]]:
|
55 |
"""
|
56 |
+
Predict probabilities for all Lionguard categories given an input text.
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
57 |
Args:
|
58 |
+
text (str): The input text to predict on.
|
59 |
+
config (Dict[str, Any]): The model configuration.
|
60 |
Returns:
|
61 |
+
Dict[str, Dict[str, Any]]: A dictionary containing prediction results for each category.
|
62 |
"""
|
63 |
+
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
|
64 |
+
embeddings = get_embeddings(device, [text], config)
|
65 |
+
X_input = np.array(embeddings, dtype=np.float32)
|
66 |
+
|
67 |
+
results = {}
|
68 |
+
for category, details in config['classifier'].items():
|
69 |
+
local_model_fp = hf_hub_download(repo_id=REPO_PATH, filename=details['model_fp'])
|
70 |
+
session = rt.InferenceSession(local_model_fp)
|
71 |
+
input_name = session.get_inputs()[0].name
|
72 |
+
outputs = session.run(None, {input_name: X_input})
|
73 |
+
|
74 |
+
if details['calibrated']:
|
75 |
+
scores = [output[1] for output in outputs[1]]
|
76 |
+
else:
|
77 |
+
scores = outputs[1].flatten()
|
78 |
+
|
79 |
+
results[category] = {
|
80 |
+
'score': float(scores[0]),
|
81 |
+
'predictions': {
|
82 |
+
'high_recall': 1 if scores[0] >= details['threshold']['high_recall'] else 0,
|
83 |
+
'balanced': 1 if scores[0] >= details['threshold']['balanced'] else 0,
|
84 |
+
'high_precision': 1 if scores[0] >= details['threshold']['high_precision'] else 0
|
85 |
+
}
|
86 |
+
}
|
87 |
+
|
88 |
+
return results
|
89 |
+
|
90 |
+
def predict_and_format(text: str, config: Dict[str, Any]) -> pd.DataFrame:
|
91 |
"""
|
92 |
+
Predict and format the results for display in the Gradio interface.
|
|
|
93 |
Args:
|
94 |
text (str): The input text to predict on.
|
95 |
+
config (Dict[str, Any]): The model configuration.
|
|
|
96 |
Returns:
|
97 |
+
pd.DataFrame: A DataFrame containing prediction results for each category.
|
98 |
"""
|
99 |
if not text.strip():
|
100 |
return None
|
101 |
|
102 |
+
results = predict(text, config)
|
103 |
+
formatted_results = []
|
104 |
+
for category, result in results.items():
|
105 |
+
formatted_results.append({
|
106 |
+
"Category": category,
|
107 |
+
"Score": f"{result['score']:.3f}",
|
108 |
+
"High Recall": result['predictions']['high_recall'],
|
109 |
+
"Balanced": result['predictions']['balanced'],
|
110 |
+
"High Precision": result['predictions']['high_precision']
|
111 |
+
})
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
112 |
|
113 |
+
return pd.DataFrame(formatted_results).sort_values("Score", ascending=False)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
114 |
|
115 |
+
def create_interface(config: Dict[str, Any]) -> gr.Interface:
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
116 |
"""
|
117 |
Create the Gradio interface for the Lionguard demo.
|
|
|
118 |
Args:
|
119 |
+
config (Dict[str, Any]): The model configuration.
|
|
|
120 |
Returns:
|
121 |
gr.Interface: The Gradio interface object.
|
122 |
"""
|
123 |
return gr.Interface(
|
124 |
+
fn=lambda text: predict_and_format(text, config),
|
125 |
inputs=gr.Textbox(lines=3, placeholder="Enter text here..."),
|
126 |
+
outputs=gr.DataFrame(label="Prediction Results"),
|
127 |
title="🦁 Lionguard Demo",
|
128 |
description="Lionguard is a Singapore-contextualized moderation classifier that can serve against unsafe LLM outputs.",
|
129 |
allow_flagging="never"
|
130 |
)
|
131 |
|
132 |
if __name__ == "__main__":
|
133 |
+
config = load_config()
|
134 |
+
iface = create_interface(config)
|
135 |
iface.launch()
|