Spaces:
Sleeping
Sleeping
Update app.py
Browse files
app.py
CHANGED
|
@@ -10,7 +10,6 @@ import gradio as gr
|
|
| 10 |
# -----------------------------
|
| 11 |
MODEL_NAME = "fakespot-ai/roberta-base-ai-text-detection-v1"
|
| 12 |
|
| 13 |
-
# Global placeholders for lazy loading (prevents startup crashes)
|
| 14 |
tokenizer = None
|
| 15 |
model = None
|
| 16 |
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
|
|
@@ -19,14 +18,14 @@ def get_model():
|
|
| 19 |
global tokenizer, model
|
| 20 |
if model is None:
|
| 21 |
tokenizer = AutoTokenizer.from_pretrained(MODEL_NAME)
|
| 22 |
-
# 2026 Check: Use bfloat16 for modern GPU efficiency
|
| 23 |
dtype = torch.bfloat16 if (device.type == "cuda" and torch.cuda.is_bf16_supported()) else torch.float32
|
| 24 |
model = AutoModelForSequenceClassification.from_pretrained(
|
| 25 |
MODEL_NAME, torch_dtype=dtype
|
| 26 |
).to(device).eval()
|
| 27 |
return tokenizer, model
|
| 28 |
|
| 29 |
-
|
|
|
|
| 30 |
|
| 31 |
# -----------------------------
|
| 32 |
# PROTECT STRUCTURE
|
|
@@ -75,14 +74,12 @@ def analyze(text):
|
|
| 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 "—", "—", "<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)
|
|
@@ -91,10 +88,8 @@ def analyze(text):
|
|
| 91 |
|
| 92 |
inputs = tok(windows, return_tensors="pt", padding=True, truncation=True, max_length=512).to(device)
|
| 93 |
logits = mod(**inputs).logits
|
| 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
|
|
@@ -103,8 +98,6 @@ def analyze(text):
|
|
| 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):
|
|
@@ -115,7 +108,6 @@ def analyze(text):
|
|
| 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:
|
|
@@ -135,7 +127,7 @@ def analyze(text):
|
|
| 135 |
|
| 136 |
highlighted_html += "</div>"
|
| 137 |
|
| 138 |
-
#
|
| 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]})
|
|
|
|
| 10 |
# -----------------------------
|
| 11 |
MODEL_NAME = "fakespot-ai/roberta-base-ai-text-detection-v1"
|
| 12 |
|
|
|
|
| 13 |
tokenizer = None
|
| 14 |
model = None
|
| 15 |
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
|
|
|
|
| 18 |
global tokenizer, model
|
| 19 |
if model is None:
|
| 20 |
tokenizer = AutoTokenizer.from_pretrained(MODEL_NAME)
|
|
|
|
| 21 |
dtype = torch.bfloat16 if (device.type == "cuda" and torch.cuda.is_bf16_supported()) else torch.float32
|
| 22 |
model = AutoModelForSequenceClassification.from_pretrained(
|
| 23 |
MODEL_NAME, torch_dtype=dtype
|
| 24 |
).to(device).eval()
|
| 25 |
return tokenizer, model
|
| 26 |
|
| 27 |
+
# Threshold set to 20% per your request
|
| 28 |
+
THRESHOLD = 0.20
|
| 29 |
|
| 30 |
# -----------------------------
|
| 31 |
# PROTECT STRUCTURE
|
|
|
|
| 74 |
return "ERROR", "0%", f"Failed to load model: {str(e)}", None
|
| 75 |
|
| 76 |
blocks = split_preserving_structure(text)
|
|
|
|
| 77 |
pure_sents_indices = [i for i, b in enumerate(blocks) if b.strip() and not b.startswith("\n")]
|
| 78 |
pure_sents = [blocks[i] for i in pure_sents_indices]
|
| 79 |
|
| 80 |
if not pure_sents:
|
| 81 |
return "—", "—", "<em>No sentences detected.</em>", None
|
| 82 |
|
|
|
|
| 83 |
windows = []
|
| 84 |
for i in range(len(pure_sents)):
|
| 85 |
start = max(0, i - 1)
|
|
|
|
| 88 |
|
| 89 |
inputs = tok(windows, return_tensors="pt", padding=True, truncation=True, max_length=512).to(device)
|
| 90 |
logits = mod(**inputs).logits
|
|
|
|
| 91 |
probs = F.softmax(logits.float(), dim=-1)[:, 1].cpu().numpy().tolist()
|
| 92 |
|
|
|
|
| 93 |
lengths = [len(s.split()) for s in pure_sents]
|
| 94 |
total_words = sum(lengths)
|
| 95 |
weighted_avg = sum(p * l for p, l in zip(probs, lengths)) / total_words if total_words > 0 else 0
|
|
|
|
| 98 |
# HTML RECONSTRUCTION
|
| 99 |
# -----------------------------
|
| 100 |
highlighted_html = "<div style='font-family: sans-serif; line-height: 1.8;'>"
|
|
|
|
|
|
|
| 101 |
prob_map = {idx: probs[i] for i, idx in enumerate(pure_sents_indices)}
|
| 102 |
|
| 103 |
for i, block in enumerate(blocks):
|
|
|
|
| 108 |
highlighted_html += block
|
| 109 |
continue
|
| 110 |
|
|
|
|
| 111 |
if i in prob_map:
|
| 112 |
score = prob_map[i]
|
| 113 |
if score < 0.35:
|
|
|
|
| 127 |
|
| 128 |
highlighted_html += "</div>"
|
| 129 |
|
| 130 |
+
# Logic Updated: > 20% triggers detection label
|
| 131 |
label = "AI Content Detected" if weighted_avg >= THRESHOLD else "0 or * AI Content Detected"
|
| 132 |
|
| 133 |
df = pd.DataFrame({"Sentence": pure_sents, "AI Confidence": [f"{p:.1%}" for p in probs]})
|