Spaces:
Running
Running
Update app.py
Browse files
app.py
CHANGED
|
@@ -6,185 +6,158 @@ import pandas as pd
|
|
| 6 |
import gradio as gr
|
| 7 |
|
| 8 |
# -----------------------------
|
| 9 |
-
# MODEL
|
| 10 |
# -----------------------------
|
| 11 |
MODEL_NAME = "fakespot-ai/roberta-base-ai-text-detection-v1"
|
| 12 |
tokenizer = AutoTokenizer.from_pretrained(MODEL_NAME)
|
| 13 |
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
|
| 14 |
dtype = torch.bfloat16 if (device.type == "cuda" and torch.cuda.is_bf16_supported()) else torch.float32
|
| 15 |
-
model = AutoModelForSequenceClassification.from_pretrained(MODEL_NAME, dtype=dtype).to(device).eval()
|
| 16 |
|
| 17 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
| 18 |
|
| 19 |
# -----------------------------
|
| 20 |
-
#
|
| 21 |
# -----------------------------
|
| 22 |
-
ABBR = [
|
| 23 |
-
"e.g", "i.e", "mr", "mrs", "ms", "dr", "prof", "vs", "etc",
|
| 24 |
-
"fig", "al", "jr", "sr", "st", "no", "vol", "pp", "mt",
|
| 25 |
-
"inc", "ltd", "co", "u.s", "u.k", "a.m", "p.m"
|
| 26 |
-
]
|
| 27 |
-
|
| 28 |
ABBR_REGEX = re.compile(r"\b(" + "|".join(map(re.escape, ABBR)) + r")\.", re.IGNORECASE)
|
| 29 |
|
| 30 |
-
|
| 31 |
def _protect(text):
|
| 32 |
text = text.replace("...", "⟨ELLIPSIS⟩")
|
| 33 |
text = re.sub(r"(?<=\d)\.(?=\d)", "⟨DECIMAL⟩", text)
|
| 34 |
text = ABBR_REGEX.sub(r"\1⟨ABBRDOT⟩", text)
|
| 35 |
return text
|
| 36 |
|
| 37 |
-
|
| 38 |
def _restore(text):
|
| 39 |
-
return (
|
| 40 |
-
text.replace("⟨ABBRDOT⟩", ".")
|
| 41 |
-
.replace("⟨DECIMAL⟩", ".")
|
| 42 |
-
.replace("⟨ELLIPSIS⟩", "...")
|
| 43 |
-
)
|
| 44 |
-
|
| 45 |
|
| 46 |
-
# -----------------------------
|
| 47 |
-
# PERFECT PARAGRAPH-PRESERVING SPLITTER
|
| 48 |
-
# -----------------------------
|
| 49 |
def split_preserving_structure(text):
|
| 50 |
-
blocks = re.split(r"(\n+)", text)
|
| 51 |
final_blocks = []
|
| 52 |
-
|
| 53 |
for block in blocks:
|
| 54 |
if block.startswith("\n"):
|
| 55 |
final_blocks.append(block)
|
| 56 |
else:
|
| 57 |
protected = _protect(block)
|
|
|
|
| 58 |
parts = re.split(r"([.?!])(\s+)", protected)
|
| 59 |
-
|
| 60 |
for i in range(0, len(parts), 3):
|
| 61 |
sentence = parts[i]
|
| 62 |
-
punct = parts[i
|
| 63 |
-
space = parts[i
|
| 64 |
-
|
| 65 |
-
|
| 66 |
-
if whole.strip():
|
| 67 |
-
final_blocks.append(_restore(whole))
|
| 68 |
if space:
|
| 69 |
final_blocks.append(space)
|
| 70 |
-
|
| 71 |
return final_blocks
|
| 72 |
|
| 73 |
-
|
| 74 |
-
def extract_sentences_only(blocks):
|
| 75 |
-
return [
|
| 76 |
-
b for b in blocks
|
| 77 |
-
if b.strip() != "" and not b.startswith("\n") and not b.isspace()
|
| 78 |
-
]
|
| 79 |
-
|
| 80 |
-
|
| 81 |
# -----------------------------
|
| 82 |
-
#
|
| 83 |
# -----------------------------
|
| 84 |
-
|
| 85 |
-
|
| 86 |
-
|
| 87 |
-
|
| 88 |
-
# -----------------------------
|
| 89 |
-
# ANALYSIS LOGIC
|
| 90 |
-
# -----------------------------
|
| 91 |
-
def analyze(text, max_len=512):
|
| 92 |
|
| 93 |
blocks = split_preserving_structure(text)
|
| 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 |
for block in blocks:
|
| 128 |
if block.startswith("\n"):
|
| 129 |
-
|
| 130 |
continue
|
| 131 |
-
|
| 132 |
if block.isspace():
|
| 133 |
-
|
| 134 |
continue
|
| 135 |
-
|
| 136 |
-
|
| 137 |
-
|
| 138 |
-
|
| 139 |
-
|
| 140 |
-
|
| 141 |
-
|
| 142 |
-
|
| 143 |
-
|
| 144 |
-
|
| 145 |
-
|
| 146 |
-
color = "#11823b"
|
| 147 |
-
elif ai_p < 0.70:
|
| 148 |
-
color = "#b8860b"
|
| 149 |
else:
|
| 150 |
-
color = "#b80d0d"
|
|
|
|
| 151 |
|
| 152 |
-
|
| 153 |
-
f"<span style='background:
|
| 154 |
-
f"
|
| 155 |
-
f"{
|
| 156 |
)
|
|
|
|
| 157 |
|
| 158 |
-
|
| 159 |
-
|
| 160 |
-
|
| 161 |
-
|
| 162 |
-
|
| 163 |
-
|
| 164 |
-
|
| 165 |
-
df = pd.DataFrame(
|
| 166 |
-
[[i + 1, s, ai_scores[i]] for i, s in enumerate(pure_sentences)],
|
| 167 |
-
columns=["#", "Sentence", "AI_Prob"]
|
| 168 |
-
)
|
| 169 |
-
|
| 170 |
-
return overall_label, overall_pct, highlighted, df
|
| 171 |
|
|
|
|
| 172 |
|
| 173 |
# -----------------------------
|
| 174 |
-
#
|
| 175 |
# -----------------------------
|
| 176 |
-
with gr.Blocks() as demo:
|
| 177 |
-
gr.Markdown("##
|
| 178 |
-
|
| 179 |
-
|
| 180 |
-
|
| 181 |
-
|
| 182 |
-
|
| 183 |
-
|
| 184 |
-
|
| 185 |
-
|
| 186 |
-
|
| 187 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 188 |
|
| 189 |
if __name__ == "__main__":
|
| 190 |
-
demo.launch()
|
|
|
|
| 6 |
import gradio as gr
|
| 7 |
|
| 8 |
# -----------------------------
|
| 9 |
+
# MODEL INITIALIZATION
|
| 10 |
# -----------------------------
|
| 11 |
MODEL_NAME = "fakespot-ai/roberta-base-ai-text-detection-v1"
|
| 12 |
tokenizer = AutoTokenizer.from_pretrained(MODEL_NAME)
|
| 13 |
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
|
| 14 |
dtype = torch.bfloat16 if (device.type == "cuda" and torch.cuda.is_bf16_supported()) else torch.float32
|
|
|
|
| 15 |
|
| 16 |
+
model = AutoModelForSequenceClassification.from_pretrained(
|
| 17 |
+
MODEL_NAME, torch_dtype=dtype
|
| 18 |
+
).to(device).eval()
|
| 19 |
+
|
| 20 |
+
THRESHOLD = 0.75 # Balanced threshold for Fakespot 2025
|
| 21 |
|
| 22 |
# -----------------------------
|
| 23 |
+
# STRUCTURAL PRESERVATION
|
| 24 |
# -----------------------------
|
| 25 |
+
ABBR = ["e.g", "i.e", "mr", "mrs", "ms", "dr", "prof", "vs", "etc", "fig", "al", "jr", "sr", "st", "inc", "ltd", "u.s", "u.k"]
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 26 |
ABBR_REGEX = re.compile(r"\b(" + "|".join(map(re.escape, ABBR)) + r")\.", re.IGNORECASE)
|
| 27 |
|
|
|
|
| 28 |
def _protect(text):
|
| 29 |
text = text.replace("...", "⟨ELLIPSIS⟩")
|
| 30 |
text = re.sub(r"(?<=\d)\.(?=\d)", "⟨DECIMAL⟩", text)
|
| 31 |
text = ABBR_REGEX.sub(r"\1⟨ABBRDOT⟩", text)
|
| 32 |
return text
|
| 33 |
|
|
|
|
| 34 |
def _restore(text):
|
| 35 |
+
return text.replace("⟨ABBRDOT⟩", ".").replace("⟨DECIMAL⟩", ".").replace("⟨ELLIPSIS⟩", "...")
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 36 |
|
|
|
|
|
|
|
|
|
|
| 37 |
def split_preserving_structure(text):
|
| 38 |
+
blocks = re.split(r"(\n+)", text)
|
| 39 |
final_blocks = []
|
|
|
|
| 40 |
for block in blocks:
|
| 41 |
if block.startswith("\n"):
|
| 42 |
final_blocks.append(block)
|
| 43 |
else:
|
| 44 |
protected = _protect(block)
|
| 45 |
+
# Split by punctuation followed by space
|
| 46 |
parts = re.split(r"([.?!])(\s+)", protected)
|
|
|
|
| 47 |
for i in range(0, len(parts), 3):
|
| 48 |
sentence = parts[i]
|
| 49 |
+
punct = parts[i+1] if i+1 < len(parts) else ""
|
| 50 |
+
space = parts[i+2] if i+2 < len(parts) else ""
|
| 51 |
+
if sentence.strip():
|
| 52 |
+
final_blocks.append(_restore(sentence + punct))
|
|
|
|
|
|
|
| 53 |
if space:
|
| 54 |
final_blocks.append(space)
|
|
|
|
| 55 |
return final_blocks
|
| 56 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 57 |
# -----------------------------
|
| 58 |
+
# ADVANCED DETECTION LOGIC
|
| 59 |
# -----------------------------
|
| 60 |
+
@torch.inference_mode()
|
| 61 |
+
def analyze(text):
|
| 62 |
+
if not text.strip():
|
| 63 |
+
return "—", "—", "<em>Please enter text...</em>", None
|
|
|
|
|
|
|
|
|
|
|
|
|
| 64 |
|
| 65 |
blocks = split_preserving_structure(text)
|
| 66 |
+
# Filter out whitespace/newlines for the model
|
| 67 |
+
pure_sents = [b for b in blocks if b.strip() and not b.startswith("\n")]
|
| 68 |
+
|
| 69 |
+
if not pure_sents:
|
| 70 |
+
return "—", "—", "<em>No sentences detected.</em>", None
|
| 71 |
+
|
| 72 |
+
# 1. Create Overlapping Context Windows (Size=3, Stride=1)
|
| 73 |
+
# This ensures every sentence is evaluated in multiple contexts
|
| 74 |
+
windows = []
|
| 75 |
+
for i in range(len(pure_sents)):
|
| 76 |
+
start = max(0, i - 1)
|
| 77 |
+
end = min(len(pure_sents), i + 2)
|
| 78 |
+
windows.append(" ".join(pure_sents[start:end]))
|
| 79 |
+
|
| 80 |
+
# 2. Batched Inference
|
| 81 |
+
inputs = tokenizer(windows, return_tensors="pt", padding=True, truncation=True, max_length=512).to(device)
|
| 82 |
+
logits = model(**inputs).logits
|
| 83 |
+
probs = F.softmax(logits, dim=-1)[:, 1].cpu().float().numpy().tolist()
|
| 84 |
+
|
| 85 |
+
# 3. Map Scores (Average scores if a sentence appeared in multiple windows)
|
| 86 |
+
# In this logic, probs[i] corresponds directly to pure_sents[i] as the 'center'
|
| 87 |
+
ai_scores = probs
|
| 88 |
+
|
| 89 |
+
# 4. Weighted Overall Score (Length matters!)
|
| 90 |
+
# AI models are more confident on longer sequences.
|
| 91 |
+
lengths = [len(s.split()) for s in pure_sents]
|
| 92 |
+
total_words = sum(lengths)
|
| 93 |
+
weighted_avg = sum(p * l for p, l in zip(ai_scores, lengths)) / total_words if total_words > 0 else 0
|
| 94 |
|
| 95 |
# -----------------------------
|
| 96 |
+
# HTML HIGHLIGHTING RECONSTRUCTION
|
| 97 |
# -----------------------------
|
| 98 |
+
highlighted_html = "<div style='font-family: sans-serif; line-height: 1.6;'>"
|
| 99 |
+
sent_ptr = 0
|
| 100 |
+
|
| 101 |
for block in blocks:
|
| 102 |
if block.startswith("\n"):
|
| 103 |
+
highlighted_html += block.replace("\n", "<br>")
|
| 104 |
continue
|
|
|
|
| 105 |
if block.isspace():
|
| 106 |
+
highlighted_html += block
|
| 107 |
continue
|
| 108 |
+
|
| 109 |
+
score = ai_scores[sent_ptr]
|
| 110 |
+
sent_ptr += 1
|
| 111 |
+
|
| 112 |
+
# Dynamic color scaling
|
| 113 |
+
if score < 0.30:
|
| 114 |
+
color = "#11823b" # Green
|
| 115 |
+
bg = "rgba(17, 130, 59, 0.1)"
|
| 116 |
+
elif score < 0.70:
|
| 117 |
+
color = "#b8860b" # Gold
|
| 118 |
+
bg = "rgba(184, 134, 11, 0.1)"
|
|
|
|
|
|
|
|
|
|
| 119 |
else:
|
| 120 |
+
color = "#b80d0d" # Red
|
| 121 |
+
bg = "rgba(184, 13, 13, 0.1)"
|
| 122 |
|
| 123 |
+
highlighted_html += (
|
| 124 |
+
f"<span style='background:{bg}; padding:2px 4px; border-radius:4px; border-bottom: 2px solid {color};' "
|
| 125 |
+
f"title='AI Probability: {score:.1%}'>"
|
| 126 |
+
f"<small style='color:{color}; font-weight:bold;'>[{score:.0%}]</small> {block}</span>"
|
| 127 |
)
|
| 128 |
+
highlighted_html += "</div>"
|
| 129 |
|
| 130 |
+
label = "🤖 Likely AI Written" if weighted_avg >= THRESHOLD else "🧒 Likely Human Written"
|
| 131 |
+
|
| 132 |
+
df = pd.DataFrame({
|
| 133 |
+
"Sentence": pure_sents,
|
| 134 |
+
"AI Score": [f"{p:.1%}" for p in ai_scores]
|
| 135 |
+
})
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 136 |
|
| 137 |
+
return label, f"{weighted_avg:.1%}", highlighted_html, df
|
| 138 |
|
| 139 |
# -----------------------------
|
| 140 |
+
# GRADIO INTERFACE
|
| 141 |
# -----------------------------
|
| 142 |
+
with gr.Blocks(theme=gr.themes.Soft()) as demo:
|
| 143 |
+
gr.Markdown("## 🕵️ AI Detector Pro")
|
| 144 |
+
gr.Markdown("Sentence-level heatmap with **Weighted Sliding Window** logic.")
|
| 145 |
+
|
| 146 |
+
with gr.Row():
|
| 147 |
+
with gr.Column(scale=3):
|
| 148 |
+
text_input = gr.Textbox(label="Paste Text", lines=12)
|
| 149 |
+
run_btn = gr.Button("Analyze Content", variant="primary")
|
| 150 |
+
with gr.Column(scale=1):
|
| 151 |
+
verdict_out = gr.Label(label="Verdict")
|
| 152 |
+
score_out = gr.Label(label="Weighted AI Score")
|
| 153 |
+
|
| 154 |
+
with gr.Tabs():
|
| 155 |
+
with gr.TabItem("Analysis Heatmap"):
|
| 156 |
+
html_out = gr.HTML()
|
| 157 |
+
with gr.TabItem("Detailed Data"):
|
| 158 |
+
table_out = gr.Dataframe(headers=["Sentence", "AI Score"], wrap=True)
|
| 159 |
+
|
| 160 |
+
run_btn.click(analyze, inputs=text_input, outputs=[verdict_out, score_out, html_out, table_out])
|
| 161 |
|
| 162 |
if __name__ == "__main__":
|
| 163 |
+
demo.launch()
|