Upload llm_functionality.py

llm_functionality.py

@@ -0,0 +1,86 @@
+from transformers import pipeline, AutoModelForSequenceClassification, AutoTokenizer
+import torch
+from typing import Literal
+
+# Initialize models
+profanity_model_name = "Dabid/abusive-tagalog-profanity-detection"
+privacy_model_name = "roberta-base"
+
+# Load profanity model
+profanity_tokenizer = AutoTokenizer.from_pretrained(profanity_model_name)
+profanity_model = AutoModelForSequenceClassification.from_pretrained(profanity_model_name)
+
+# Load privacy model
+privacy_tokenizer = AutoTokenizer.from_pretrained(privacy_model_name)
+privacy_model = AutoModelForSequenceClassification.from_pretrained(privacy_model_name)
+
+# Use GPU if available
+device = 0 if torch.cuda.is_available() else -1
+
+# Create classifiers
+profanity_classifier = pipeline(
+    "text-classification",
+    model=profanity_model,
+    tokenizer=profanity_tokenizer,
+    device=device,
+    framework="pt"
+)
+
+privacy_classifier = pipeline(
+    "text-classification",
+    model=privacy_model,
+    tokenizer=privacy_tokenizer,
+    device=device,
+    framework="pt"
+)
+
+
+PROMPT_TEMPLATES = {
+    "Profanity Detection": {
+        "system": """Detect profanity in debt collection conversations. Flag if the text contains:
+        1. Explicit swear words (e.g., "damn", "hell", "crap").
+        2. Threats ("Pay or else").
+        3. Derogatory terms ("idiot", "fraud").
+        4. Aggressive tone ("Listen up!").
+
+        Examples:
+        - Text: "Pay your damn bill!" → Label: 1
+        - Text: "Can we discuss payment options?" → Label: 0
+
+        Text: "{text}"
+        Label:""",
+        "label_map": {"LABEL_0": 0, "LABEL_1": 1}
+    },
+
+    "Privacy Violation": {
+        "system": """Flag privacy violations if:
+            1. Financial details (balance, account info) are shared **before** verifying:
+            - Full address, DOB, or SSN (last 4 digits).
+            2. Verification is skipped or unconfirmed by the customer.
+
+            Examples:
+            - No Violation: Agent asks for DOB first, then shares balance. → Label: 0
+            - Violation: Agent says, "You owe $300" without verification. → Label: 1
+
+            Text: "{text}"
+            Label:""",
+            "label_map": {"LABEL_0": 0, "LABEL_1": 1}
+    }
+}
+
+
+
+def analyze_text(text: str, analysis_type: Literal["Profanity Detection", "Privacy Violation"]):
+    """Analyses the given text using Profanity and Compliance LLMs."""
+    template = PROMPT_TEMPLATES[analysis_type]
+    
+    if analysis_type == "Profanity Detection":
+        # Use the Tagalog-English profanity model directly
+        result = profanity_classifier(text, truncation=True, max_length=512)
+        return 1 if result[0]['label'] == 'LABEL_1' else 0
+    else:
+        # Use RoBERTa with prompt engineering for privacy checks
+        prompt = template["system"].format(text=text)
+        result = privacy_classifier(prompt, truncation=True, max_length=512)
+        return template["label_map"][result[0]['label']]
+