Spaces:
Sleeping
Sleeping
Upload 3 files
Browse files
model_inference/__init__.py
ADDED
File without changes
|
model_inference/gaurdrails_manager.py
ADDED
@@ -0,0 +1,64 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
from endpoints.api_models import OutputGuardrailsConfig , LLMResponse
|
2 |
+
from model_inference.groundedness_checker import GroundednessChecker
|
3 |
+
import re
|
4 |
+
|
5 |
+
#
|
6 |
+
groundedness_checker = GroundednessChecker(model_path="./grounding_detector")
|
7 |
+
|
8 |
+
# A simple result class to hold individual check outcomes.
|
9 |
+
class Result:
|
10 |
+
def __init__(self):
|
11 |
+
self.details = {}
|
12 |
+
|
13 |
+
def add(self, rule_name: str, passed: bool):
|
14 |
+
self.details[rule_name] = passed
|
15 |
+
|
16 |
+
def grounded(self) -> bool:
|
17 |
+
# The response is considered "grounded" if all enabled rules pass.
|
18 |
+
return all(self.details.values())
|
19 |
+
|
20 |
+
class ContextualGroundednessCheck:
|
21 |
+
name = "Contextual Groundedness"
|
22 |
+
|
23 |
+
def check(self,llm_response:LLMResponse) -> bool:
|
24 |
+
groundedness_check = groundedness_checker.check(llm_response.question, llm_response.answer, llm_response.context)
|
25 |
+
print(groundedness_check)
|
26 |
+
return groundedness_check['is_grounded']
|
27 |
+
|
28 |
+
|
29 |
+
class ToxicityRule:
|
30 |
+
|
31 |
+
name = "Toxicity"
|
32 |
+
|
33 |
+
def check(self, llm_response:LLMResponse) -> bool:
|
34 |
+
|
35 |
+
no_toxicity = True
|
36 |
+
matched = re.search(r"(hate|kill|suicide|selfharm)", llm_response.answer, re.IGNORECASE)
|
37 |
+
|
38 |
+
if matched:
|
39 |
+
no_toxicity = False
|
40 |
+
|
41 |
+
return no_toxicity
|
42 |
+
|
43 |
+
|
44 |
+
# Manager class to load and execute the enabled guardrail rules.
|
45 |
+
class GuardrailsManager:
|
46 |
+
def __init__(self, config: OutputGuardrailsConfig):
|
47 |
+
self.config = config
|
48 |
+
self.rules = self.load_rules()
|
49 |
+
|
50 |
+
def load_rules(self):
|
51 |
+
rules = []
|
52 |
+
if self.config.contextual_grounding:
|
53 |
+
rules.append(ContextualGroundednessCheck())
|
54 |
+
if self.config.toxicity:
|
55 |
+
rules.append(ToxicityRule())
|
56 |
+
# Add additional rules based on configuration here.
|
57 |
+
return rules
|
58 |
+
|
59 |
+
def check(self, llm_response: LLMResponse) -> Result:
|
60 |
+
result = Result()
|
61 |
+
for rule in self.rules:
|
62 |
+
rule_result = rule.check(llm_response)
|
63 |
+
result.add(rule.name, rule_result)
|
64 |
+
return result
|
model_inference/groundedness_checker.py
ADDED
@@ -0,0 +1,64 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
from transformers import AutoTokenizer, AutoModelForSequenceClassification
|
2 |
+
import torch
|
3 |
+
|
4 |
+
class GroundednessChecker:
|
5 |
+
def __init__(self, model_path="./grounding_detector"):
|
6 |
+
self.tokenizer = AutoTokenizer.from_pretrained(model_path)
|
7 |
+
self.model = AutoModelForSequenceClassification.from_pretrained(model_path)
|
8 |
+
self.device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
|
9 |
+
self.model.to(self.device)
|
10 |
+
|
11 |
+
def check(self, question: str, answer: str, context: str) -> dict:
|
12 |
+
"""Check if answer is grounded in context"""
|
13 |
+
inputs = self.tokenizer(
|
14 |
+
question,
|
15 |
+
answer + " [SEP] " + context,
|
16 |
+
padding=True,
|
17 |
+
truncation=True,
|
18 |
+
max_length=512,
|
19 |
+
return_tensors="pt"
|
20 |
+
).to(self.device)
|
21 |
+
|
22 |
+
with torch.no_grad():
|
23 |
+
outputs = self.model(**inputs)
|
24 |
+
|
25 |
+
probs = torch.nn.functional.softmax(outputs.logits, dim=-1)
|
26 |
+
return {
|
27 |
+
"is_grounded": bool(torch.argmax(probs)),
|
28 |
+
"confidence": probs[0][1].item(),
|
29 |
+
"details": {
|
30 |
+
"question": question,
|
31 |
+
"answer": answer,
|
32 |
+
"context_snippet": context[:200] + "..." if len(context) > 200 else context
|
33 |
+
}
|
34 |
+
}
|
35 |
+
|
36 |
+
# Usage Example
|
37 |
+
if __name__ == "__main__":
|
38 |
+
# Initialize checker
|
39 |
+
checker = GroundednessChecker()
|
40 |
+
|
41 |
+
# Example from banking PDS
|
42 |
+
context = """
|
43 |
+
Premium Savings Account Terms:
|
44 |
+
- Annual Percentage Yield (APY): 4.25%
|
45 |
+
- Minimum opening deposit: $1,000
|
46 |
+
- Monthly maintenance fee: $5 (waived if daily balance >= $1,000)
|
47 |
+
- Maximum withdrawals: 6 per month
|
48 |
+
"""
|
49 |
+
|
50 |
+
# Grounded example
|
51 |
+
grounded_result = checker.check(
|
52 |
+
question="What is the minimum opening deposit?",
|
53 |
+
answer="$1,000",
|
54 |
+
context=context
|
55 |
+
)
|
56 |
+
print("Grounded Result:", grounded_result)
|
57 |
+
|
58 |
+
# Ungrounded example
|
59 |
+
ungrounded_result = checker.check(
|
60 |
+
question="What is the monthly maintenance fee?",
|
61 |
+
answer="$10 monthly charge",
|
62 |
+
context=context
|
63 |
+
)
|
64 |
+
print("Ungrounded Result:", ungrounded_result)
|