Spaces:
Running
Running
Update app.py
Browse files
app.py
CHANGED
|
@@ -8,7 +8,7 @@ import gradio as gr
|
|
| 8 |
# -----------------------------
|
| 9 |
# MODEL INITIALIZATION
|
| 10 |
# -----------------------------
|
| 11 |
-
MODEL_NAME = "
|
| 12 |
tokenizer = None
|
| 13 |
model = None
|
| 14 |
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
|
|
@@ -16,21 +16,20 @@ device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
|
|
| 16 |
def get_model():
|
| 17 |
global tokenizer, model
|
| 18 |
if model is None:
|
| 19 |
-
print(f"Loading
|
| 20 |
-
tokenizer = AutoTokenizer.from_pretrained(MODEL_NAME)
|
| 21 |
-
dtype = torch.float32
|
| 22 |
-
if device.type == "cuda" and torch.cuda.is_bf16_supported():
|
| 23 |
-
dtype = torch.bfloat16
|
| 24 |
model = AutoModelForSequenceClassification.from_pretrained(
|
| 25 |
-
MODEL_NAME,
|
|
|
|
|
|
|
| 26 |
).to(device).eval()
|
| 27 |
return tokenizer, model
|
| 28 |
|
| 29 |
-
# UPDATED
|
| 30 |
-
THRESHOLD = 0.
|
| 31 |
|
| 32 |
# -----------------------------
|
| 33 |
-
#
|
| 34 |
# -----------------------------
|
| 35 |
ABBR = ["e.g", "i.e", "mr", "mrs", "ms", "dr", "prof", "vs", "etc", "fig", "al", "jr", "sr", "st", "inc", "ltd", "u.s", "u.k"]
|
| 36 |
ABBR_REGEX = re.compile(r"\b(" + "|".join(map(re.escape, ABBR)) + r")\.", re.IGNORECASE)
|
|
@@ -73,6 +72,7 @@ def analyze(text):
|
|
| 73 |
return "—", "—", "<em>Please enter text...</em>", None
|
| 74 |
|
| 75 |
word_count = len(text.split())
|
|
|
|
| 76 |
if word_count < 300:
|
| 77 |
warning_msg = f"⚠️ <b>Insufficient Text:</b> Your input has {word_count} words. Please enter at least 300 words for an accurate analysis."
|
| 78 |
return "Too Short", "N/A", f"<div style='color: #b80d0d; padding: 20px; border: 1px solid #b80d0d; border-radius: 8px;'>{warning_msg}</div>", None
|
|
@@ -96,8 +96,8 @@ def analyze(text):
|
|
| 96 |
windows.append(" ".join(pure_sents[start:end]))
|
| 97 |
|
| 98 |
inputs = tok(windows, return_tensors="pt", padding=True, truncation=True, max_length=512).to(device)
|
| 99 |
-
|
| 100 |
-
probs =
|
| 101 |
|
| 102 |
lengths = [len(s.split()) for s in pure_sents]
|
| 103 |
total_words = sum(lengths)
|
|
@@ -116,8 +116,7 @@ def analyze(text):
|
|
| 116 |
|
| 117 |
if i in prob_map:
|
| 118 |
score = prob_map[i]
|
| 119 |
-
|
| 120 |
-
# Logic: Red for > 0.81, Green for everything else (<= 0.81)
|
| 121 |
if score >= THRESHOLD:
|
| 122 |
color, bg = "#b80d0d", "rgba(184, 13, 13, 0.15)" # RED
|
| 123 |
else:
|
|
@@ -125,38 +124,33 @@ def analyze(text):
|
|
| 125 |
|
| 126 |
highlighted_html += (
|
| 127 |
f"<span style='background:{bg}; padding:2px 4px; border-radius:4px; border-bottom: 2px solid {color};' "
|
| 128 |
-
f"title='
|
| 129 |
-
f"<b style='color:{color}; font-size: 0.8em;'>[{score:.
|
| 130 |
)
|
| 131 |
else:
|
| 132 |
highlighted_html += block
|
| 133 |
highlighted_html += "</div>"
|
| 134 |
|
| 135 |
-
|
| 136 |
-
|
| 137 |
-
label = f"{weighted_avg:.0%} AI Content Detected"
|
| 138 |
-
display_score = f"{weighted_avg:.1%}"
|
| 139 |
-
else:
|
| 140 |
-
label = "0 or * AI Content Detected"
|
| 141 |
-
display_score = "*"
|
| 142 |
|
| 143 |
-
df = pd.DataFrame({"Sentence": pure_sents, "AI Confidence": [f"{p:.
|
| 144 |
return label, display_score, highlighted_html, df
|
| 145 |
|
| 146 |
# -----------------------------
|
| 147 |
# GRADIO INTERFACE
|
| 148 |
# -----------------------------
|
| 149 |
with gr.Blocks(theme=gr.themes.Soft()) as demo:
|
| 150 |
-
gr.Markdown("## 🕵️ AI Detector Pro")
|
| 151 |
-
gr.Markdown(f"
|
| 152 |
|
| 153 |
with gr.Row():
|
| 154 |
with gr.Column(scale=3):
|
| 155 |
-
text_input = gr.Textbox(label="Paste Text", lines=12, placeholder="
|
| 156 |
run_btn = gr.Button("Analyze", variant="primary")
|
| 157 |
with gr.Column(scale=1):
|
| 158 |
-
verdict_out = gr.Label(label="Verdict")
|
| 159 |
-
score_out = gr.Label(label="Weighted
|
| 160 |
|
| 161 |
with gr.Tabs():
|
| 162 |
with gr.TabItem("Visual Heatmap"):
|
|
|
|
| 8 |
# -----------------------------
|
| 9 |
# MODEL INITIALIZATION
|
| 10 |
# -----------------------------
|
| 11 |
+
MODEL_NAME = "desklib/ai-text-detector-v1.01"
|
| 12 |
tokenizer = None
|
| 13 |
model = None
|
| 14 |
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
|
|
|
|
| 16 |
def get_model():
|
| 17 |
global tokenizer, model
|
| 18 |
if model is None:
|
| 19 |
+
print(f"Loading Specialized Model: {MODEL_NAME} on {device}")
|
| 20 |
+
tokenizer = AutoTokenizer.from_pretrained(MODEL_NAME, use_fast=False)
|
|
|
|
|
|
|
|
|
|
| 21 |
model = AutoModelForSequenceClassification.from_pretrained(
|
| 22 |
+
MODEL_NAME,
|
| 23 |
+
num_labels=1,
|
| 24 |
+
ignore_mismatched_sizes=True
|
| 25 |
).to(device).eval()
|
| 26 |
return tokenizer, model
|
| 27 |
|
| 28 |
+
# UPDATED: Threshold changed from 0.81 to 0.59
|
| 29 |
+
THRESHOLD = 0.59
|
| 30 |
|
| 31 |
# -----------------------------
|
| 32 |
+
# UTILITIES
|
| 33 |
# -----------------------------
|
| 34 |
ABBR = ["e.g", "i.e", "mr", "mrs", "ms", "dr", "prof", "vs", "etc", "fig", "al", "jr", "sr", "st", "inc", "ltd", "u.s", "u.k"]
|
| 35 |
ABBR_REGEX = re.compile(r"\b(" + "|".join(map(re.escape, ABBR)) + r")\.", re.IGNORECASE)
|
|
|
|
| 72 |
return "—", "—", "<em>Please enter text...</em>", None
|
| 73 |
|
| 74 |
word_count = len(text.split())
|
| 75 |
+
# Word count remains at 300
|
| 76 |
if word_count < 300:
|
| 77 |
warning_msg = f"⚠️ <b>Insufficient Text:</b> Your input has {word_count} words. Please enter at least 300 words for an accurate analysis."
|
| 78 |
return "Too Short", "N/A", f"<div style='color: #b80d0d; padding: 20px; border: 1px solid #b80d0d; border-radius: 8px;'>{warning_msg}</div>", None
|
|
|
|
| 96 |
windows.append(" ".join(pure_sents[start:end]))
|
| 97 |
|
| 98 |
inputs = tok(windows, return_tensors="pt", padding=True, truncation=True, max_length=512).to(device)
|
| 99 |
+
output = mod(**inputs)
|
| 100 |
+
probs = torch.sigmoid(output.logits).cpu().numpy().flatten().tolist()
|
| 101 |
|
| 102 |
lengths = [len(s.split()) for s in pure_sents]
|
| 103 |
total_words = sum(lengths)
|
|
|
|
| 116 |
|
| 117 |
if i in prob_map:
|
| 118 |
score = prob_map[i]
|
| 119 |
+
# Logic now uses the 0.59 threshold
|
|
|
|
| 120 |
if score >= THRESHOLD:
|
| 121 |
color, bg = "#b80d0d", "rgba(184, 13, 13, 0.15)" # RED
|
| 122 |
else:
|
|
|
|
| 124 |
|
| 125 |
highlighted_html += (
|
| 126 |
f"<span style='background:{bg}; padding:2px 4px; border-radius:4px; border-bottom: 2px solid {color};' "
|
| 127 |
+
f"title='Raw Model Score: {score:.4f}'>"
|
| 128 |
+
f"<b style='color:{color}; font-size: 0.8em;'>[{score:.1%}]</b> {block}</span>"
|
| 129 |
)
|
| 130 |
else:
|
| 131 |
highlighted_html += block
|
| 132 |
highlighted_html += "</div>"
|
| 133 |
|
| 134 |
+
label = f"{weighted_avg:.1%} AI Probability"
|
| 135 |
+
display_score = f"{weighted_avg:.2%}"
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 136 |
|
| 137 |
+
df = pd.DataFrame({"Sentence": pure_sents, "AI Confidence": [f"{p:.2%}" for p in probs]})
|
| 138 |
return label, display_score, highlighted_html, df
|
| 139 |
|
| 140 |
# -----------------------------
|
| 141 |
# GRADIO INTERFACE
|
| 142 |
# -----------------------------
|
| 143 |
with gr.Blocks(theme=gr.themes.Soft()) as demo:
|
| 144 |
+
gr.Markdown("## 🕵️ AI Detector Pro: Raw Mode")
|
| 145 |
+
gr.Markdown(f"Direct model output from **{MODEL_NAME}**. Visual highlight now triggers at **{THRESHOLD*100:.0f}%**.")
|
| 146 |
|
| 147 |
with gr.Row():
|
| 148 |
with gr.Column(scale=3):
|
| 149 |
+
text_input = gr.Textbox(label="Paste Text", lines=12, placeholder="Min 300 words...")
|
| 150 |
run_btn = gr.Button("Analyze", variant="primary")
|
| 151 |
with gr.Column(scale=1):
|
| 152 |
+
verdict_out = gr.Label(label="Model Verdict (Raw)")
|
| 153 |
+
score_out = gr.Label(label="Exact Weighted Probability")
|
| 154 |
|
| 155 |
with gr.Tabs():
|
| 156 |
with gr.TabItem("Visual Heatmap"):
|