TetherSST / app.py
SamanthaStorm's picture
Update app.py
d4d856f verified
raw
history blame
6.52 kB
import gradio as gr
import torch
from transformers import pipeline as hf_pipeline, AutoModelForSequenceClassification, AutoTokenizer
# ——— 1) Emotion Pipeline ————————————————————————————————————————————————
emotion_pipeline = hf_pipeline(
"text-classification",
model="j-hartmann/emotion-english-distilroberta-base",
top_k=None,
truncation=True
)
def get_emotion_profile(text):
"""
Returns a dict of { emotion_label: score } for the input text.
"""
results = emotion_pipeline(text)
# some pipelines return [[…]]
if isinstance(results, list) and isinstance(results[0], list):
results = results[0]
return {r["label"].lower(): round(r["score"], 3) for r in results}
# ——— 2) Abuse-patterns Model ——————————————————————————————————————————————
model_name = "SamanthaStorm/tether-multilabel-v3"
model = AutoModelForSequenceClassification.from_pretrained(model_name)
tokenizer = AutoTokenizer.from_pretrained(model_name, use_fast=False)
LABELS = [
"blame shifting", "contradictory statements", "control", "dismissiveness",
"gaslighting", "guilt tripping", "insults", "obscure language",
"projection", "recovery phase", "threat"
]
THRESHOLDS = {
"blame shifting": 0.28, "contradictory statements": 0.27, "control": 0.08, "dismissiveness": 0.32,
"gaslighting": 0.27, "guilt tripping": 0.31, "insults": 0.10, "obscure language": 0.55,
"projection": 0.09, "recovery phase": 0.33, "threat": 0.15
}
# ——— 3) Emotional-tone tagging (no abuse_score / DARVO) —————————————————————————————
def get_tone_tag(emotion_profile, patterns):
anger = emotion_profile.get("anger", 0)
disgust = emotion_profile.get("disgust", 0)
sadness = emotion_profile.get("sadness", 0)
joy = emotion_profile.get("joy", 0)
neutral = emotion_profile.get("neutral", 0)
fear = emotion_profile.get("fear", 0)
# 1) Vulnerable: sadness high + recovery-phase
if sadness > 0.4 and "recovery phase" in patterns:
return "vulnerable"
# 2) Supportive: joy very high + no other patterns (or only recovery-phase)
if joy > 0.5 and (not patterns or patterns == ["recovery phase"]):
return "supportive"
# 3) Confrontational: anger/disgust high + aggressive patterns
if (anger + disgust) > 0.5 and any(p in patterns for p in ["insults", "control", "threat"]):
return "confrontational"
# 4) Manipulative: neutral high + classic manipulation patterns
if neutral > 0.4 and any(p in patterns for p in ["gaslighting", "dismissiveness", "projection", "guilt tripping", "blame shifting"]):
return "manipulative"
# 5) Feigned Warmth: joy high but manipulative patterns present
if joy > 0.5 and any(p in patterns for p in ["gaslighting", "dismissiveness", "projection", "guilt tripping", "blame shifting"]):
return "feigned warmth"
# 6) Defensive: anger high + contradictory statements
if anger > 0.4 and "contradictory statements" in patterns:
return "defensive"
# 7) Neutral: pure neutral dominates all
if neutral > max(anger, disgust, sadness, joy, fear):
return "neutral"
return None
# ——— 3) Single-message analysis ——————————————————————————————————————————————
def analyze_message(text):
"""
Runs emotion profiling, and the abuse-pattern classifier.
Returns a dict with:
- emotion_profile: { emotion: score }
- active_patterns: [ labels above their threshold ]
"""
emotion_profile = get_emotion_profile(text)
# get raw model scores
toks = tokenizer(text, return_tensors="pt", truncation=True, padding=True)
with torch.no_grad():
logits = model(**toks).logits.squeeze(0)
scores = torch.sigmoid(logits).cpu().numpy()
# pick up all labels whose score >= threshold
active = [lab for lab, sc in zip(LABELS, scores) if sc >= THRESHOLDS[lab]]
return {
"emotion_profile": emotion_profile,
"active_patterns": active,
"tone_tag": tone_tag
}
# ——— 5) Composite wrapper (handles .txt or image + text boxes) ——————————————————————
def analyze_composite(uploaded_file, *texts):
outputs = []
if uploaded_file is not None:
raw = uploaded_file.read()
name = uploaded_file.name.lower()
if name.endswith((".png", ".jpg", ".jpeg", ".tiff", ".bmp", ".gif")):
img = Image.open(io.BytesIO(raw))
content = pytesseract.image_to_string(img)
else:
try:
content = raw.decode("utf-8")
except UnicodeDecodeError:
content = raw.decode("latin-1")
r = analyze_message(content)
outputs.append(
"── Uploaded File ──\n"
f"Emotion Profile : {r['emotion_profile']}\n"
f"Active Patterns : {r['active_patterns']}\n"
f"Emotional Tone : {r['tone_tag']}\n"
)
for idx, txt in enumerate(texts, start=1):
if not txt:
continue
r = analyze_message(txt)
outputs.append(
f"── Message {idx} ──\n"
f"Emotion Profile : {r['emotion_profile']}\n"
f"Active Patterns : {r['active_patterns']}\n"
f"Emotional Tone : {r['tone_tag']}\n"
)
return "\n".join(outputs) if outputs else "Please enter at least one message."
# ——— 6) Gradio interface ————————————————————————————————————————————————
message_inputs = [gr.Textbox(label=f"Message {i+1}") for i in range(3)]
iface = gr.Interface(
fn=analyze_composite,
inputs=[
gr.File(file_types=[".txt", ".png", ".jpg", ".jpeg"],
label="Upload text or image")
] + message_inputs,
outputs=gr.Textbox(label="Analysis"),
title="Tether Analyzer (with Tone Tags)",
description="Extracts motifs, emotions, patterns—and now an emotional tone tag—no abuse score or DARVO."
)
if __name__ == "__main__":
iface.launch()