app.py

import gradio as gr
import torch
import numpy as np
from transformers import AutoModelForSequenceClassification, AutoTokenizer
from transformers import RobertaForSequenceClassification, RobertaTokenizer

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

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”,
“contradictory_statements”, “manipulation”, “deflection”, “insults”, “obscure_formal”, “recovery_phase”, “non_abusive”,
“suicidal_threat”, “physical_threat”, “extreme_control”
]

THRESHOLDS = {
“gaslighting”: 0.25, “mockery”: 0.15, “dismissiveness”: 0.30, “control”: 0.43, “guilt_tripping”: 0.19,
“apology_baiting”: 0.45, “blame_shifting”: 0.23, “projection”: 0.50, “contradictory_statements”: 0.25,
“manipulation”: 0.25, “deflection”: 0.30, “insults”: 0.34, “obscure_formal”: 0.25, “recovery_phase”: 0.25,
“non_abusive”: 2.0, “suicidal_threat”: 0.45, “physical_threat”: 0.02, “extreme_control”: 0.36
}

PATTERN_LABELS = LABELS[:15]
DANGER_LABELS = LABELS[15:18]

EXPLANATIONS = {
“gaslighting”: “Gaslighting involves making someone question their own reality or perceptions…”,
“blame_shifting”: “Blame-shifting is when one person redirects the responsibility…”,
“projection”: “Projection involves accusing the victim of behaviors the abuser exhibits.”,
“dismissiveness”: “Dismissiveness is belittling or disregarding another person’s feelings.”,
“mockery”: “Mockery ridicules someone in a hurtful, humiliating way.”,
“recovery_phase”: “Recovery phase dismisses someone’s emotional healing process.”,
“insults”: “Insults are derogatory remarks aimed at degrading someone.”,
“apology_baiting”: “Apology-baiting manipulates victims into apologizing for abuser’s behavior.”,
“deflection”: “Deflection avoids accountability by redirecting blame.”,
“control”: “Control restricts autonomy through manipulation or coercion.”,
“extreme_control”: “Extreme control dominates decisions and behaviors entirely.”,
“physical_threat”: “Physical threats signal risk of bodily harm.”,
“suicidal_threat”: “Suicidal threats manipulate others using self-harm threats.”,
“guilt_tripping”: “Guilt-tripping uses guilt to manipulate someone’s actions.”,
“manipulation”: “Manipulation deceives to influence or control outcomes.”,
“non_abusive”: “Non-abusive language is respectful and free of coercion.”,
“obscure_formal”: “Obscure/formal language manipulates through confusion or superiority.”
}

def custom_sentiment(text):
inputs = sentiment_tokenizer(text, return_tensors=“pt”, truncation=True, padding=True)
with torch.no_grad():
outputs = sentiment_model(**inputs)
probs = torch.nn.functional.softmax(outputs.logits, dim=1)
label_idx = torch.argmax(probs).item()
label_map = {0: “supportive”, 1: “undermining”}
label = label_map[label_idx]
score = probs[0][label_idx].item()
return {“label”: label, “score”: score}

def calculate_abuse_level(scores, thresholds):
triggered_scores = [score for label, score in zip(LABELS, scores) if score > thresholds[label]]
return round(np.mean(triggered_scores) * 100, 2) if triggered_scores else 0.0

def interpret_abuse_level(score):
if score > 80: return “Extreme / High Risk”
elif score > 60: return “Severe / Harmful Pattern Present”
elif score > 40: return “Likely Abuse”
elif score > 20: return “Mild Concern”
return “Very Low / Likely Safe”

def analyze_messages(input_text, risk_flags):
input_text = input_text.strip()
if not input_text:
return “Please enter a message for analysis.”

sentiment = custom_sentiment(input_text)
sentiment_label = sentiment['label']
sentiment_score = sentiment['score']

adjusted_thresholds = {k: v * 0.8 for k, v in THRESHOLDS.items()} if sentiment_label == "undermining" else THRESHOLDS.copy()

inputs = tokenizer(input_text, return_tensors="pt", truncation=True, padding=True)
with torch.no_grad():
    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]))
danger_flag_count = sum(score > adjusted_thresholds[label] for label, score in zip(DANGER_LABELS, scores[15:18]))

contextual_flags = risk_flags if risk_flags else []
if len(contextual_flags) >= 2:
    danger_flag_count += 1

critical_flags = ["They've threatened harm", "They monitor/follow me", "I feel unsafe when alone with them"]
high_risk_context = any(flag in contextual_flags for flag in critical_flags)

non_abusive_score = scores[LABELS.index('non_abusive')]
if non_abusive_score > adjusted_thresholds['non_abusive']:
    return "This message is classified as non-abusive."

abuse_level = calculate_abuse_level(scores, adjusted_thresholds)
abuse_description = interpret_abuse_level(abuse_level)

if danger_flag_count >= 2:
    resources = "Immediate assistance recommended. Please seek professional help or contact emergency services."
else:
    resources = "For more information on abuse patterns, consider reaching out to support groups or professional counselors."

scored_patterns = [(label, scores[LABELS.index(label)]) for label in PATTERN_LABELS if label != "non_abusive"]
scored_dangers = [(label, scores[LABELS.index(label)]) for label in DANGER_LABELS]

top_patterns = sorted(
    [(label, score) for label, score in scored_patterns if score > adjusted_thresholds[label]],
    key=lambda x: x[1], reverse=True
)[:2]

top_dangers = sorted(
    [(label, score) for label, score in scored_dangers if score > adjusted_thresholds[label]],
    key=lambda x: x[1], reverse=True
)

highlighted_danger = top_dangers[0] if top_dangers else None

top_pattern_explanations = "\n".join([
    f"\u2022 {label.replace('_', ' ').title()}: {EXPLANATIONS.get(label, 'No explanation available.')}"
    for label, _ in top_patterns
])

if highlighted_danger and highlighted_danger[0] not in [label for label, _ in top_patterns]:
    danger_label = highlighted_danger[0]
    top_pattern_explanations += f"\n\u2022 {danger_label.replace('_', ' ').title()}: {EXPLANATIONS.get(danger_label, 'No explanation available.')} (Danger Pattern)"

result = (
    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"
    "Resources: " + resources + "\n\n"
    f"Sentiment: {sentiment_label.title()} (Confidence: {sentiment_score*100:.2f}%)"
)

if contextual_flags:
    result += "\n\n⚠️ You indicated the following:\n" + "\n".join([f"• {flag}" for flag in contextual_flags])
if high_risk_context:
    result += "\n\n🚨 These responses suggest a high-risk situation. Consider seeking immediate help or safety planning resources."

return result

iface = gr.Interface(
fn=analyze_messages,
inputs=[
gr.Textbox(lines=10, placeholder=“Enter message here…”),
gr.CheckboxGroup(label=“Do any of these apply to your situation?”, choices=[
“They’ve threatened harm”, “They isolate me”, “I’ve changed my behavior out of fear”,
“They monitor/follow me”, “I feel unsafe when alone with them”
])
],
outputs=[gr.Textbox(label=“Analysis Result”)],
title=“Abuse Pattern Detector”
)

if name == “main”:
iface.launch()