Spaces:
Running
Running
Update app.py
Browse files
app.py
CHANGED
|
@@ -69,19 +69,20 @@ def split_preserving_structure(text):
|
|
| 69 |
def analyze(text):
|
| 70 |
if not text.strip():
|
| 71 |
return "β", "β", "<em>Please enter text...</em>", None
|
| 72 |
-
|
| 73 |
try:
|
| 74 |
tok, mod = get_model()
|
| 75 |
except Exception as e:
|
| 76 |
return "ERROR", "0%", f"Failed to load model: {str(e)}", None
|
| 77 |
|
| 78 |
blocks = split_preserving_structure(text)
|
| 79 |
-
|
| 80 |
-
|
|
|
|
|
|
|
| 81 |
if not pure_sents:
|
| 82 |
return "β", "β", "<em>No sentences detected.</em>", None
|
| 83 |
|
| 84 |
-
# Sliding Window (Size 3, Stride 1)
|
| 85 |
windows = []
|
| 86 |
for i in range(len(pure_sents)):
|
| 87 |
start = max(0, i - 1)
|
|
@@ -93,7 +94,7 @@ def analyze(text):
|
|
| 93 |
# Ensure stable math with .float() before softmax
|
| 94 |
probs = F.softmax(logits.float(), dim=-1)[:, 1].cpu().numpy().tolist()
|
| 95 |
|
| 96 |
-
# Weighted Scoring
|
| 97 |
lengths = [len(s.split()) for s in pure_sents]
|
| 98 |
total_words = sum(lengths)
|
| 99 |
weighted_avg = sum(p * l for p, l in zip(probs, lengths)) / total_words if total_words > 0 else 0
|
|
@@ -102,36 +103,42 @@ def analyze(text):
|
|
| 102 |
# HTML RECONSTRUCTION
|
| 103 |
# -----------------------------
|
| 104 |
highlighted_html = "<div style='font-family: sans-serif; line-height: 1.8;'>"
|
| 105 |
-
|
| 106 |
-
|
|
|
|
|
|
|
|
|
|
| 107 |
if block.startswith("\n"):
|
| 108 |
highlighted_html += block.replace("\n", "<br>")
|
| 109 |
continue
|
| 110 |
if block.isspace():
|
| 111 |
highlighted_html += block
|
| 112 |
continue
|
| 113 |
-
|
| 114 |
-
score = probs[sent_ptr]
|
| 115 |
-
sent_ptr += 1
|
| 116 |
|
| 117 |
-
#
|
| 118 |
-
if
|
| 119 |
-
|
| 120 |
-
|
| 121 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 122 |
else:
|
| 123 |
-
|
| 124 |
|
| 125 |
-
highlighted_html += (
|
| 126 |
-
f"<span style='background:{bg}; padding:2px 4px; border-radius:4px; border-bottom: 2px solid {color};' "
|
| 127 |
-
f"title='AI Probability: {score:.1%}'>"
|
| 128 |
-
f"<b style='color:{color}; font-size: 0.8em;'>[{score:.0%}]</b> {block}</span>"
|
| 129 |
-
)
|
| 130 |
highlighted_html += "</div>"
|
| 131 |
-
|
|
|
|
| 132 |
label = "AI Content Detected" if weighted_avg >= THRESHOLD else "0 or * AI Content Detected"
|
|
|
|
| 133 |
df = pd.DataFrame({"Sentence": pure_sents, "AI Confidence": [f"{p:.1%}" for p in probs]})
|
| 134 |
-
|
| 135 |
return label, f"{weighted_avg:.1%}", highlighted_html, df
|
| 136 |
|
| 137 |
# -----------------------------
|
|
@@ -140,7 +147,7 @@ def analyze(text):
|
|
| 140 |
with gr.Blocks(theme=gr.themes.Soft()) as demo:
|
| 141 |
gr.Markdown("## π΅οΈ AI Detector Pro")
|
| 142 |
gr.Markdown("Sentence-level analysis with weighted context windows.")
|
| 143 |
-
|
| 144 |
with gr.Row():
|
| 145 |
with gr.Column(scale=3):
|
| 146 |
text_input = gr.Textbox(label="Paste Text", lines=12)
|
|
@@ -148,13 +155,13 @@ with gr.Blocks(theme=gr.themes.Soft()) as demo:
|
|
| 148 |
with gr.Column(scale=1):
|
| 149 |
verdict_out = gr.Label(label="Verdict")
|
| 150 |
score_out = gr.Label(label="Weighted AI Score")
|
| 151 |
-
|
| 152 |
with gr.Tabs():
|
| 153 |
with gr.TabItem("Visual Heatmap"):
|
| 154 |
html_out = gr.HTML()
|
| 155 |
with gr.TabItem("Raw Data Breakdown"):
|
| 156 |
table_out = gr.Dataframe(headers=["Sentence", "AI Confidence"], wrap=True)
|
| 157 |
-
|
| 158 |
run_btn.click(analyze, inputs=text_input, outputs=[verdict_out, score_out, html_out, table_out])
|
| 159 |
|
| 160 |
if __name__ == "__main__":
|
|
|
|
| 69 |
def analyze(text):
|
| 70 |
if not text.strip():
|
| 71 |
return "β", "β", "<em>Please enter text...</em>", None
|
|
|
|
| 72 |
try:
|
| 73 |
tok, mod = get_model()
|
| 74 |
except Exception as e:
|
| 75 |
return "ERROR", "0%", f"Failed to load model: {str(e)}", None
|
| 76 |
|
| 77 |
blocks = split_preserving_structure(text)
|
| 78 |
+
# Track indices of actual sentences to align probabilities correctly
|
| 79 |
+
pure_sents_indices = [i for i, b in enumerate(blocks) if b.strip() and not b.startswith("\n")]
|
| 80 |
+
pure_sents = [blocks[i] for i in pure_sents_indices]
|
| 81 |
+
|
| 82 |
if not pure_sents:
|
| 83 |
return "β", "β", "<em>No sentences detected.</em>", None
|
| 84 |
|
| 85 |
+
# Sliding Window (Size 3, Stride 1)
|
| 86 |
windows = []
|
| 87 |
for i in range(len(pure_sents)):
|
| 88 |
start = max(0, i - 1)
|
|
|
|
| 94 |
# Ensure stable math with .float() before softmax
|
| 95 |
probs = F.softmax(logits.float(), dim=-1)[:, 1].cpu().numpy().tolist()
|
| 96 |
|
| 97 |
+
# Weighted Scoring
|
| 98 |
lengths = [len(s.split()) for s in pure_sents]
|
| 99 |
total_words = sum(lengths)
|
| 100 |
weighted_avg = sum(p * l for p, l in zip(probs, lengths)) / total_words if total_words > 0 else 0
|
|
|
|
| 103 |
# HTML RECONSTRUCTION
|
| 104 |
# -----------------------------
|
| 105 |
highlighted_html = "<div style='font-family: sans-serif; line-height: 1.8;'>"
|
| 106 |
+
|
| 107 |
+
# Map probabilities to the original block indices
|
| 108 |
+
prob_map = {idx: probs[i] for i, idx in enumerate(pure_sents_indices)}
|
| 109 |
+
|
| 110 |
+
for i, block in enumerate(blocks):
|
| 111 |
if block.startswith("\n"):
|
| 112 |
highlighted_html += block.replace("\n", "<br>")
|
| 113 |
continue
|
| 114 |
if block.isspace():
|
| 115 |
highlighted_html += block
|
| 116 |
continue
|
|
|
|
|
|
|
|
|
|
| 117 |
|
| 118 |
+
# Only highlight if the block was actually a sentence we analyzed
|
| 119 |
+
if i in prob_map:
|
| 120 |
+
score = prob_map[i]
|
| 121 |
+
if score < 0.35:
|
| 122 |
+
color, bg = "#11823b", "rgba(17, 130, 59, 0.15)"
|
| 123 |
+
elif score < 0.70:
|
| 124 |
+
color, bg = "#b8860b", "rgba(184, 134, 11, 0.15)"
|
| 125 |
+
else:
|
| 126 |
+
color, bg = "#b80d0d", "rgba(184, 13, 13, 0.15)"
|
| 127 |
+
|
| 128 |
+
highlighted_html += (
|
| 129 |
+
f"<span style='background:{bg}; padding:2px 4px; border-radius:4px; border-bottom: 2px solid {color};' "
|
| 130 |
+
f"title='AI Probability: {score:.1%}'>"
|
| 131 |
+
f"<b style='color:{color}; font-size: 0.8em;'>[{score:.0%}]</b> {block}</span>"
|
| 132 |
+
)
|
| 133 |
else:
|
| 134 |
+
highlighted_html += block
|
| 135 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 136 |
highlighted_html += "</div>"
|
| 137 |
+
|
| 138 |
+
# Restored original labels exactly as requested
|
| 139 |
label = "AI Content Detected" if weighted_avg >= THRESHOLD else "0 or * AI Content Detected"
|
| 140 |
+
|
| 141 |
df = pd.DataFrame({"Sentence": pure_sents, "AI Confidence": [f"{p:.1%}" for p in probs]})
|
|
|
|
| 142 |
return label, f"{weighted_avg:.1%}", highlighted_html, df
|
| 143 |
|
| 144 |
# -----------------------------
|
|
|
|
| 147 |
with gr.Blocks(theme=gr.themes.Soft()) as demo:
|
| 148 |
gr.Markdown("## π΅οΈ AI Detector Pro")
|
| 149 |
gr.Markdown("Sentence-level analysis with weighted context windows.")
|
| 150 |
+
|
| 151 |
with gr.Row():
|
| 152 |
with gr.Column(scale=3):
|
| 153 |
text_input = gr.Textbox(label="Paste Text", lines=12)
|
|
|
|
| 155 |
with gr.Column(scale=1):
|
| 156 |
verdict_out = gr.Label(label="Verdict")
|
| 157 |
score_out = gr.Label(label="Weighted AI Score")
|
| 158 |
+
|
| 159 |
with gr.Tabs():
|
| 160 |
with gr.TabItem("Visual Heatmap"):
|
| 161 |
html_out = gr.HTML()
|
| 162 |
with gr.TabItem("Raw Data Breakdown"):
|
| 163 |
table_out = gr.Dataframe(headers=["Sentence", "AI Confidence"], wrap=True)
|
| 164 |
+
|
| 165 |
run_btn.click(analyze, inputs=text_input, outputs=[verdict_out, score_out, html_out, table_out])
|
| 166 |
|
| 167 |
if __name__ == "__main__":
|