Update app.py

app.py

@@ -4,15 +4,14 @@ import numpy as np
 from transformers import AutoModelForSequenceClassification, AutoTokenizer
 from transformers import RobertaForSequenceClassification, RobertaTokenizer
 
-# Load fine-tuned sentiment model (DistilBERT)
-sentiment_model_name = "SamanthaStorm/tether-sentiment"
-sentiment_model = AutoModelForSequenceClassification.from_pretrained(sentiment_model_name)
-sentiment_tokenizer = AutoTokenizer.from_pretrained(sentiment_model_name)
+# Load custom fine-tuned sentiment model
+sentiment_model = AutoModelForSequenceClassification.from_pretrained("SamanthaStorm/tether-sentiment")
+sentiment_tokenizer = AutoTokenizer.from_pretrained("SamanthaStorm/tether-sentiment")
 
-# Load abuse pattern model (RoBERTa)
-abuse_model_name = "SamanthaStorm/abuse-pattern-detector-v2"
-abuse_model = RobertaForSequenceClassification.from_pretrained(abuse_model_name)
-abuse_tokenizer = RobertaTokenizer.from_pretrained(abuse_model_name)
+# Load abuse pattern model
+model_name = "SamanthaStorm/abuse-pattern-detector-v2"
+model = RobertaForSequenceClassification.from_pretrained(model_name, trust_remote_code=True)
+tokenizer = RobertaTokenizer.from_pretrained(model_name, trust_remote_code=True)
 
 LABELS = [
     "gaslighting", "mockery", "dismissiveness", "control", "guilt_tripping", "apology_baiting", "blame_shifting", "projection",
@@ -80,11 +79,11 @@ def analyze_messages(input_text, risk_flags):
     sentiment_label = sentiment['label']
     sentiment_score = sentiment['score']
 
-    adjusted_thresholds = {k: v * 0.8 for k, v in THRESHOLDS.items()} if sentiment_label == "NEGATIVE" else THRESHOLDS.copy()
+    adjusted_thresholds = {k: v * 0.8 for k, v in THRESHOLDS.items()} if sentiment_label == "undermining" else THRESHOLDS.copy()
 
-    inputs = abuse_tokenizer(input_text, return_tensors="pt", truncation=True, padding=True)
+    inputs = tokenizer(input_text, return_tensors="pt", truncation=True, padding=True)
     with torch.no_grad():
-        outputs = abuse_model(**inputs)
+        outputs = model(**inputs)
     scores = torch.sigmoid(outputs.logits.squeeze(0)).numpy()
 
     pattern_count = sum(score > adjusted_thresholds[label] for label, score in zip(PATTERN_LABELS, scores[:15]))
@@ -101,7 +100,7 @@ def analyze_messages(input_text, risk_flags):
     if non_abusive_score > adjusted_thresholds['non_abusive']:
         return "This message is classified as non-abusive."
 
-    abuse_level = calculate_abuse_level(scores, THRESHOLDS)
+    abuse_level = calculate_abuse_level(scores, adjusted_thresholds)
     abuse_description = interpret_abuse_level(abuse_level)
 
     if danger_flag_count >= 2:
@@ -121,8 +120,8 @@ def analyze_messages(input_text, risk_flags):
         f"Abuse Risk Score: {abuse_level}% – {abuse_description}\n\n"
         f"Most Likely Patterns:\n{top_pattern_explanations}\n\n"
         f"⚠️ Critical Danger Flags Detected: {danger_flag_count} of 3\n"
-        f"Resources: {resources}\n\n"
-        f"Sentiment: {sentiment_label} (Confidence: {sentiment_score*100:.2f}%)"
+        "Resources: " + resources + "\n\n"
+        f"Sentiment: {sentiment_label.title()} (Confidence: {sentiment_score*100:.2f}%)"
     )
 
     if contextual_flags: