Spaces:
Sleeping
Sleeping
File size: 5,436 Bytes
d08010b | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 | import warnings
warnings.filterwarnings("ignore")
import joblib
import numpy as np
import torch
from transformers import AutoTokenizer, AutoModelForSequenceClassification
# ============================================================
# Load TF-IDF models (trained by us on 728K samples)
# ============================================================
vectorizer = joblib.load("text_vectorizer.pkl")
lr_model = joblib.load("text_model_lr.pkl")
sgd_model = joblib.load("text_model_sgd.pkl")
# ============================================================
# Load GLYPH — DeBERTa-v3 pretrained detector
# 98.85% accuracy, covers GPT-2 through GPT-4 (14 AI families)
# Used as benchmark comparison model
# ============================================================
GLYPH_MODEL = "ogmatrixllm/glyph-v1.1"
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
print("Loading GLYPH model...")
try:
# IMPORTANT: use_fast=False required — fast tokenizer has
# a confirmed regression in transformers>=4.47 for DeBERTa-v3
glyph_tokenizer = AutoTokenizer.from_pretrained(GLYPH_MODEL, use_fast=False)
glyph_model = AutoModelForSequenceClassification.from_pretrained(GLYPH_MODEL)
glyph_model.to(device)
glyph_model.eval()
GLYPH_AVAILABLE = True
print("✓ GLYPH loaded")
print(glyph_model.config.id2label)
print("LABEL_0 = HUMAN , LABEL_1= AI")
except Exception as e:
print(f"✗ GLYPH not available: {e}")
GLYPH_AVAILABLE = False
# ============================================================
# Individual predictors
# ============================================================
def predict_lr(vec) -> dict:
proba = lr_model.predict_proba(vec)[0]
ai_prob = float(proba[1]) # index 1 = AI
return {
"label": "AI-generated" if ai_prob >= 0.5 else "Human-written",
"confidence": round(ai_prob, 4),
"ai_prob": round(ai_prob, 4)
}
def predict_sgd(vec) -> dict:
if hasattr(sgd_model, "predict_proba"):
proba = sgd_model.predict_proba(vec)[0]
ai_prob = float(proba[1])
else:
df = float(sgd_model.decision_function(vec)[0])
ai_prob = 1 / (1 + np.exp(-df))
return {
"label": "AI-generated" if ai_prob >= 0.5 else "Human-written",
"confidence": round(ai_prob, 4),
"ai_prob": round(ai_prob, 4)
}
def predict_glyph(text: str) -> dict | None:
"""
GLYPH — DeBERTa-v3-base fine-tuned detector.
98.85% accuracy across 14 AI model families (GPT-2 to GPT-4).
LABEL_0 = human, LABEL_1 = AI-generated.
"""
if not GLYPH_AVAILABLE:
return None
try:
inputs = glyph_tokenizer(
text,
return_tensors="pt",
truncation=True,
max_length=512 # trained at 512, longer texts truncated
).to(device)
with torch.no_grad():
logits = glyph_model(**inputs).logits
probs = torch.softmax(logits, dim=-1)
p_human, p_ai = probs[0].tolist()
ai_prob = round(float(p_ai), 4)
label = "AI-generated" if p_ai > 0.5 else "Human-written"
return {
"label": label,
"confidence": round(max(p_human, p_ai), 4),
"ai_prob": ai_prob
}
except Exception as e:
print(f"GLYPH error: {e}")
return None
# ============================================================
# Main predict function
# ============================================================
def predict_text(text: str) -> dict:
vec = vectorizer.transform([text])
# Run all models
lr_result = predict_lr(vec)
sgd_result = predict_sgd(vec)
glyph_result = predict_glyph(text)
# Weighted average
# GLYPH gets higher weight — 98.85% accuracy vs ~91% for TF-IDF models
scores = [lr_result["ai_prob"], sgd_result["ai_prob"]]
weights = [1.0, 1.0]
if glyph_result:
scores.append(glyph_result["ai_prob"])
weights.append(2.0) # GLYPH gets 2x weight — significantly better model
avg_ai_score = round(
sum(s * w for s, w in zip(scores, weights)) / sum(weights), 4
)
final_label = "AI-generated" if avg_ai_score >= 0.5 else "Human-written"
confidence = avg_ai_score if avg_ai_score >= 0.5 else 1 - avg_ai_score
# Warning logic
all_labels = [lr_result["label"], sgd_result["label"]]
if glyph_result:
all_labels.append(glyph_result["label"])
warning = None
if 0.4 <= avg_ai_score <= 0.6:
warning = "Borderline prediction — low confidence. Consider this result as indicative only."
elif len(set(all_labels)) > 1:
warning = "Models disagree on this text. GLYPH result is weighted higher as it covers more AI model families."
# Short text warning (GLYPH docs: <50 words has reduced F1)
if len(text.split()) < 50:
warning = (warning or "") + " Short text detected — accuracy may be reduced. Provide more text for reliable detection."
warning = warning.strip()
return {
"lr": lr_result,
"sgd": sgd_result,
"glyph": glyph_result, # None if unavailable
"final": {
"label": final_label,
"confidence": round(float(confidence), 4),
"ai_score": avg_ai_score
},
"warning": warning
}
|