Tether / app.py
SamanthaStorm's picture
Update app.py
fd6c90c verified
raw
history blame
10.1 kB
import gradio as gr
import torch
import numpy as np
from transformers import AutoTokenizer, AutoModelForSeq2SeqLM
from transformers import RobertaForSequenceClassification, RobertaTokenizer
from motif_tagging import detect_motifs
import re
# --- Sentiment Model: T5-based Emotion Classifier ---
sentiment_tokenizer = AutoTokenizer.from_pretrained("mrm8488/t5-base-finetuned-emotion")
sentiment_model = AutoModelForSeq2SeqLM.from_pretrained("mrm8488/t5-base-finetuned-emotion")
EMOTION_TO_SENTIMENT = {
"joy": "supportive",
"love": "supportive",
"surprise": "supportive",
"neutral": "supportive",
"sadness": "undermining",
"anger": "undermining",
"fear": "undermining",
"disgust": "undermining",
"shame": "undermining",
"guilt": "undermining"
}
# --- Abuse Detection Model ---
model_name = "SamanthaStorm/autotrain-jlpi4-mllvp"
model = RobertaForSequenceClassification.from_pretrained(model_name, trust_remote_code=True)
tokenizer = RobertaTokenizer.from_pretrained(model_name, trust_remote_code=True)
LABELS = [
"blame shifting", "contradictory statements", "control", "dismissiveness",
"gaslighting", "guilt tripping", "insults", "obscure language",
"projection", "recovery phase", "threat"
]
THRESHOLDS = {
"blame shifting": 0.3,
"contradictory statements": 0.32,
"control": 0.48,
"dismissiveness": 0.45,
"gaslighting": 0.30,
"guilt tripping": 0.20,
"insults": 0.34,
"obscure language": 0.25,
"projection": 0.35,
"recovery phase": 0.25,
"threat": 0.25
}
PATTERN_WEIGHTS = {
"gaslighting": 1.3,
"control": 1.2,
"dismissiveness": 0.8,
"blame shifting": 0.8,
"contradictory statements": 0.75
}
EXPLANATIONS = {
"blame shifting": "Blame-shifting is when one person redirects responsibility onto someone else to avoid accountability.",
"contradictory statements": "Contradictory statements confuse the listener by flipping positions or denying previous claims.",
"control": "Control restricts another person’s autonomy through coercion, manipulation, or threats.",
"dismissiveness": "Dismissiveness is belittling or disregarding another person’s feelings, needs, or opinions.",
"gaslighting": "Gaslighting involves making someone question their own reality, memory, or perceptions.",
"guilt tripping": "Guilt-tripping uses guilt to manipulate someone’s actions or decisions.",
"insults": "Insults are derogatory or demeaning remarks meant to shame, belittle, or hurt someone.",
"obscure language": "Obscure language manipulates through complexity, vagueness, or superiority to confuse the other person.",
"projection": "Projection accuses someone else of the very behaviors or intentions the speaker is exhibiting.",
"recovery phase": "Recovery phase statements attempt to soothe or reset tension without acknowledging harm or change.",
"threat": "Threats use fear of harm (physical, emotional, or relational) to control or intimidate someone."
}
RISK_SNIPPETS = {
"low": (
"🟢 Risk Level: Low",
"The language patterns here do not strongly indicate abuse.",
"Continue to check in with yourself and notice how you feel in response to repeated patterns."
),
"moderate": (
"⚠️ Risk Level: Moderate to High",
"This language includes control, guilt, or reversal tactics.",
"These patterns often lead to emotional confusion and reduced self-trust. Document these messages or talk with someone safe."
),
"high": (
"🛑 Risk Level: High",
"Language includes threats or coercive control, which are strong indicators of escalation.",
"Consider creating a safety plan or contacting a support line. Trust your sense of unease."
)
}
def generate_risk_snippet(abuse_score, top_label):
if abuse_score >= 85:
risk_level = "high"
elif abuse_score >= 60:
risk_level = "moderate"
else:
risk_level = "low"
title, summary, advice = RISK_SNIPPETS[risk_level]
return f"\n\n{title}\n{summary} (Pattern: **{top_label}**)\n💡 {advice}"
DARVO_PATTERNS = {
"blame shifting", "projection", "dismissiveness", "guilt tripping", "contradictory statements"
}
DARVO_MOTIFS = [
"i guess i’m the bad guy", "after everything i’ve done", "you always twist everything",
"so now it’s all my fault", "i’m the villain", "i’m always wrong", "you never listen",
"you’re attacking me", "i’m done trying", "i’m the only one who cares"
]
def detect_contradiction(message):
contradiction_phrases = [
(r"\b(i love you).{0,15}(i hate you|you ruin everything)", re.IGNORECASE),
(r"\b(i’m sorry).{0,15}(but you|if you hadn’t)", re.IGNORECASE),
(r"\b(i’m trying).{0,15}(you never|why do you)", re.IGNORECASE),
(r"\b(do what you want).{0,15}(you’ll regret it|i always give everything)", re.IGNORECASE),
(r"\b(i don’t care).{0,15}(you never think of me)", re.IGNORECASE),
(r"\b(i guess i’m just).{0,15}(the bad guy|worthless|never enough)", re.IGNORECASE)
]
return any(re.search(pattern, message, flags) for pattern, flags in contradiction_phrases)
def calculate_darvo_score(patterns, sentiment_before, sentiment_after, motifs_found, contradiction_flag=False):
pattern_hits = len([p.lower() for p in patterns if p.lower() in DARVO_PATTERNS])
pattern_score = pattern_hits / len(DARVO_PATTERNS)
sentiment_shift_score = max(0.0, sentiment_after - sentiment_before)
motif_hits = len([m.lower() for m in motifs_found if m.lower() in DARVO_MOTIFS])
motif_score = motif_hits / len(DARVO_MOTIFS)
contradiction_score = 1.0 if contradiction_flag else 0.0
darvo_score = (
0.3 * pattern_score +
0.3 * sentiment_shift_score +
0.25 * motif_score +
0.15 * contradiction_score
)
return round(min(darvo_score, 1.0), 3)
ESCALATION_QUESTIONS = [
("Partner has access to firearms or weapons", 4),
("Partner threatened to kill you", 3),
("Partner threatened you with a weapon", 3),
("Partner has ever choked you, even if you considered it consensual at the time", 4),
("Partner injured or threatened your pet(s)", 3),
("Partner has broken your things, punched or kicked walls, or thrown things ", 2),
("Partner forced or coerced you into unwanted sexual acts", 3),
("Partner threatened to take away your children", 2),
("Violence has increased in frequency or severity", 3),
("Partner monitors your calls/GPS/social media", 2)
]
def analyze_single_message(text, thresholds, motif_flags):
motif_hits, matched_phrases = detect_motifs(text)
# Sentiment Analysis
input_ids = sentiment_tokenizer(f"emotion: {text}", return_tensors="pt").input_ids
with torch.no_grad():
outputs = sentiment_model.generate(input_ids)
emotion = sentiment_tokenizer.decode(outputs[0], skip_special_tokens=True).strip().lower()
sentiment = EMOTION_TO_SENTIMENT.get(emotion, "undermining")
sentiment_score = 0.5 if sentiment == "undermining" else 0.0
# Contradiction Check
contradiction_flag = detect_contradiction(text)
# Motifs
motifs = [phrase for _, phrase in matched_phrases]
# Model Prediction
inputs = tokenizer(text, return_tensors="pt", truncation=True, padding=True)
with torch.no_grad():
outputs = model(**inputs)
scores = torch.sigmoid(outputs.logits.squeeze(0)).numpy()
threshold_labels = [label for label, score in zip(LABELS, scores) if score > thresholds[label]]
top_patterns = sorted([(label, score) for label, score in zip(LABELS, scores)], key=lambda x: x[1], reverse=True)[:2]
pattern_labels = threshold_labels + [label for label, _ in matched_phrases]
darvo_score = calculate_darvo_score(pattern_labels, 0.0, sentiment_score, motifs, contradiction_flag)
return (
np.mean([score for _, score in top_patterns]) * 100,
threshold_labels,
top_patterns,
darvo_score,
{"label": sentiment, "emotion": emotion}
)
def analyze_composite(msg1, msg2, msg3, *answers_and_none):
responses = answers_and_none[:len(ESCALATION_QUESTIONS)]
none_selected = answers_and_none[-1]
escalation_score = 0 if none_selected else sum(w for (_, w), a in zip(ESCALATION_QUESTIONS, responses) if a)
escalation_level = "High" if escalation_score >= 16 else "Moderate" if escalation_score >= 8 else "Low"
messages = [msg1, msg2, msg3]
active = [m for m in messages if m.strip()]
if not active:
return "Please enter at least one message."
results = [analyze_single_message(m, THRESHOLDS.copy(), []) for m in active]
abuse_scores = [r[0] for r in results]
darvo_scores = [r[3] for r in results]
top_pattern = max({label for r in results for label in r[2]}, key=lambda l: abuse_scores[0])
composite_abuse = int(round(sum(abuse_scores)/len(abuse_scores)))
avg_darvo = round(sum(darvo_scores)/len(darvo_scores), 3)
out = f"Abuse Intensity: {composite_abuse}%\n"
out += f"Escalation Potential: {escalation_level} ({escalation_score}/{sum(w for _,w in ESCALATION_QUESTIONS)})"
out += generate_risk_snippet(composite_abuse, top_pattern[0])
if avg_darvo > 0.25:
level = "moderate" if avg_darvo < 0.65 else "high"
out += f"\n\nDARVO Score: {avg_darvo} → This indicates a **{level} likelihood** of narrative reversal (DARVO), where the speaker may be denying, attacking, or reversing blame."
return out
textbox_inputs = [
gr.Textbox(label="Message 1"),
gr.Textbox(label="Message 2"),
gr.Textbox(label="Message 3")
]
quiz_boxes = [gr.Checkbox(label=q) for q, _ in ESCALATION_QUESTIONS]
none_box = gr.Checkbox(label="None of the above")
iface = gr.Interface(
fn=analyze_composite,
inputs=textbox_inputs + quiz_boxes + [none_box],
outputs=gr.Textbox(label="Results"),
title="Abuse Pattern Detector + Escalation Quiz",
allow_flagging="manual"
)
if __name__ == "__main__":
iface.launch()