AD232025's picture
Update model.py
7dd775e verified
import os
import re
from typing import Dict, List
from huggingface_hub import InferenceClient
# =========================================================
# HUGGING FACE INFERENCE CLIENT
# =========================================================
HF_API_TOKEN = os.getenv("HF_API_TOKEN") # optional, set in HF Space secrets
if HF_API_TOKEN:
client = InferenceClient(token=HF_API_TOKEN)
else:
client = InferenceClient() # anonymous for public models (rate-limited)
# Model IDs
TOX_MODEL_ID = "unitary/toxic-bert"
OFF_MODEL_ID = "cardiffnlp/twitter-roberta-base-offensive"
EMO_MODEL_ID = "j-hartmann/emotion-english-distilroberta-base"
SENT_MODEL_ID = "distilbert-base-uncased-finetuned-sst-2-english"
# =========================================================
# RULE KEYWORDS / PATTERNS
# =========================================================
AGGRESSION_KEYWORDS = [
"stupid", "idiot", "dumb", "incompetent", "useless",
"trash", "garbage", "worthless", "pathetic", "clown",
"moron", "failure", "shut up", "hate you"
]
THREAT_PHRASES = [
"you will regret", "there will be consequences", "watch your back",
"this is your last warning", "i'm coming for you",
"or else", "i'll ruin you", "i'll make you pay",
"i am gonna hurt you", "i'm going to hurt you",
"im gonna hurt you", # <-- added for your exact example
]
PROFANITY = [
"fuck", "shit", "bitch", "asshole", "bastard",
"motherfucker", "prick", "dickhead"
]
POLITE_KEYWORDS = [
"please", "thank you", "thanks", "would you mind",
"if possible", "kindly", "when you have a chance",
"if you don't mind"
]
FRIENDLY_KEYWORDS = [
"awesome", "amazing", "great job", "fantastic",
"love this", "appreciate you", "good vibes",
"wonderful", "you're the best", "you are the best",
]
SARCASM_PATTERNS = [
r"yeah right",
r"sure you did",
r"great job (idiot|genius)",
r"nice work (moron|buddy)",
r"well done.*not",
r"nice job.*not",
]
# Generic threat regex: “gonna/going to/will hurt you”
THREAT_REGEX = re.compile(r"\b(gonna|going to|will)\s+hurt you\b")
# =========================================================
# HF INFERENCE HELPERS
# =========================================================
def _safe_text_classification(model_id: str, text: str) -> List[Dict]:
"""
Wrapper around HF Inference API text classification.
Returns a list of dicts like:
[
{"label": "POSITIVE", "score": 0.95},
...
]
or [] on error.
"""
try:
out = client.text_classification(text, model=model_id)
# Some clients may return a single dict; normalize to list
if isinstance(out, dict):
return [out]
return out or []
except Exception as e:
print(f"[WARN] HF Inference error for {model_id}: {e}")
return []
def _get_sentiment(text: str):
"""
Returns (pos, neg) based on distilbert sentiment.
"""
results = _safe_text_classification(SENT_MODEL_ID, text)
pos = 0.5
neg = 0.5
if results:
scores = {r["label"].upper(): float(r["score"]) for r in results}
# typical labels: POSITIVE / NEGATIVE
if "POSITIVE" in scores:
pos = scores["POSITIVE"]
neg = 1.0 - pos
elif "NEGATIVE" in scores:
neg = scores["NEGATIVE"]
pos = 1.0 - neg
return pos, neg
def _get_toxicity(text: str) -> float:
"""
Return a toxicity-like score in [0, 1].
For unitary/toxic-bert, we consider any 'toxic-like' label as signal.
"""
results = _safe_text_classification(TOX_MODEL_ID, text)
if not results:
return 0.0
toxic_score = 0.0
for r in results:
label = r["label"].lower()
if any(key in label for key in ["toxic", "obscene", "insult", "hate", "threat"]):
toxic_score = max(toxic_score, float(r["score"]))
return toxic_score
def _get_offensive(text: str) -> float:
"""
Return an offensive score in [0, 1].
For cardiffnlp/twitter-roberta-base-offensive, look for OFFENSE-like labels.
"""
results = _safe_text_classification(OFF_MODEL_ID, text)
if not results:
return 0.0
off_score = 0.0
for r in results:
label = r["label"].lower()
if "offense" in label or "offensive" in label:
off_score = max(off_score, float(r["score"]))
return off_score
def _get_emotions(text: str):
"""
Returns a dict like {"anger": 0.3, "joy": 0.6}.
"""
results = _safe_text_classification(EMO_MODEL_ID, text)
if not results:
return {"anger": 0.0, "joy": 0.0}
emo = {}
for r in results:
emo[r["label"].lower()] = float(r["score"])
anger = emo.get("anger", 0.0)
joy = emo.get("joy", 0.0)
return {"anger": anger, "joy": joy}
# =========================================================
# MAIN CLASSIFIER (STRICT OPTION A)
# =========================================================
def classify_tone_rich(text: str):
lowered = text.lower()
explanation = []
# --- Model signals ---
pos, neg = _get_sentiment(text)
tox_score = _get_toxicity(text)
off_score = _get_offensive(text)
emo = _get_emotions(text)
anger = emo.get("anger", 0.0)
joy = emo.get("joy", 0.0)
explanation.append(f"Sentiment pos={pos:.2f}, neg={neg:.2f}")
explanation.append(f"Toxicity={tox_score:.2f}, Offensive={off_score:.2f}")
explanation.append(f"Emotion anger={anger:.2f}, joy={joy:.2f}")
# --- Rule flags ---
has_insult = any(w in lowered for w in AGGRESSION_KEYWORDS)
# THREATS: list OR generic regex
has_threat_phrase = any(p in lowered for p in THREAT_PHRASES)
has_threat_regex = bool(THREAT_REGEX.search(lowered))
has_threat = has_threat_phrase or has_threat_regex
has_profanity = any(bad in lowered for bad in PROFANITY)
has_polite = any(w in lowered for w in POLITE_KEYWORDS)
has_friendly = any(w in lowered for w in FRIENDLY_KEYWORDS)
has_sarcasm = any(re.search(p, lowered) for p in SARCASM_PATTERNS)
if has_insult:
explanation.append("Detected explicit insult keyword.")
if has_threat_phrase:
explanation.append("Detected explicit threat phrase.")
if has_threat_regex:
explanation.append("Matched generic threat pattern (gonna/going to/will hurt you).")
if has_profanity:
explanation.append("Detected profanity.")
if has_polite:
explanation.append("Detected polite phrasing.")
if has_friendly:
explanation.append("Detected friendly / appreciative wording.")
if has_sarcasm:
explanation.append("Matched a sarcasm pattern.")
# =====================================================
# STRICT AGGRESSIVE RULES
# =====================================================
# 1) Threats override everything
if has_threat:
return {
"label": "Aggressive",
"confidence": 95,
"severity": 95,
"threat_score": 95,
"politeness_score": 0,
"friendly_score": 0,
"has_threat": True,
"has_profanity": has_profanity,
"has_sarcasm": has_sarcasm,
"explanation": explanation,
}
# 2) Profanity → aggressive
if has_profanity:
sev = max(85, int((tox_score + off_score) / 2 * 100))
return {
"label": "Aggressive",
"confidence": 90,
"severity": sev,
"threat_score": int(tox_score * 100),
"politeness_score": 0,
"friendly_score": 0,
"has_threat": has_threat,
"has_profanity": True,
"has_sarcasm": has_sarcasm,
"explanation": explanation,
}
# 3) Direct insults → aggressive
if has_insult:
sev = max(80, int((tox_score + off_score) / 2 * 100))
return {
"label": "Aggressive",
"confidence": 88,
"severity": sev,
"threat_score": int(tox_score * 100),
"politeness_score": 0,
"friendly_score": 0,
"has_threat": has_threat,
"has_profanity": has_profanity,
"has_sarcasm": has_sarcasm,
"explanation": explanation,
}
# 4) Sarcasm + negative sentiment → aggressive
if has_sarcasm and neg > 0.55:
return {
"label": "Aggressive",
"confidence": 85,
"severity": 85,
"threat_score": int(tox_score * 100),
"politeness_score": 0,
"friendly_score": 0,
"has_threat": has_threat,
"has_profanity": has_profanity,
"has_sarcasm": True,
"explanation": explanation,
}
# 5) High anger + toxicity
if anger + tox_score > 1.1:
return {
"label": "Aggressive",
"confidence": 80,
"severity": 80,
"threat_score": int(tox_score * 100),
"politeness_score": 0,
"friendly_score": 0,
"has_threat": has_threat,
"has_profanity": has_profanity,
"has_sarcasm": has_sarcasm,
"explanation": explanation,
}
# =====================================================
# POSITIVE LABELS – FRIENDLY / POLITE
# =====================================================
if has_friendly and pos > 0.60:
return {
"label": "Friendly",
"confidence": int(pos * 100),
"severity": 0,
"threat_score": int(tox_score * 100),
"politeness_score": int(pos * 100),
"friendly_score": int(pos * 100),
"has_threat": has_threat,
"has_profanity": has_profanity,
"has_sarcasm": has_sarcasm,
"explanation": explanation,
}
if has_polite and pos > 0.50:
return {
"label": "Polite",
"confidence": int(pos * 100),
"severity": 0,
"threat_score": int(tox_score * 100),
"politeness_score": int(pos * 100),
"friendly_score": 0,
"has_threat": has_threat,
"has_profanity": has_profanity,
"has_sarcasm": has_sarcasm,
"explanation": explanation,
}
# =====================================================
# NEUTRAL FALLBACK
# =====================================================
return {
"label": "Neutral",
"confidence": int((1 - neg) * 100),
"severity": 0,
"threat_score": int(tox_score * 100),
"politeness_score": int(pos * 100),
"friendly_score": int(pos * 100),
"has_threat": has_threat,
"has_profanity": has_profanity,
"has_sarcasm": has_sarcasm,
"explanation": explanation,
}
# Optional wrapper for backwards compatibility
def classify_tone(text: str):
r = classify_tone_rich(text)
aggressive_prob = r["severity"] / 100.0
positive_prob = r["friendly_score"] / 100.0
return r["label"], r["confidence"], aggressive_prob, positive_prob