Spaces:
Sleeping
Sleeping
Update app.py
Browse files
app.py
CHANGED
|
@@ -17,9 +17,11 @@ device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
|
|
| 17 |
def get_model():
|
| 18 |
global tokenizer, model
|
| 19 |
if model is None:
|
|
|
|
|
|
|
| 20 |
tokenizer = AutoTokenizer.from_pretrained(MODEL_NAME)
|
| 21 |
|
| 22 |
-
#
|
| 23 |
dtype = torch.float32
|
| 24 |
if device.type == "cuda" and torch.cuda.is_bf16_supported():
|
| 25 |
dtype = torch.bfloat16
|
|
@@ -29,7 +31,6 @@ def get_model():
|
|
| 29 |
).to(device).eval()
|
| 30 |
return tokenizer, model
|
| 31 |
|
| 32 |
-
# Only affects the content label as requested
|
| 33 |
THRESHOLD = 0.20
|
| 34 |
|
| 35 |
# -----------------------------
|
|
@@ -67,7 +68,7 @@ def split_preserving_structure(text):
|
|
| 67 |
return final_blocks
|
| 68 |
|
| 69 |
# -----------------------------
|
| 70 |
-
#
|
| 71 |
# -----------------------------
|
| 72 |
@torch.inference_mode()
|
| 73 |
def analyze(text):
|
|
@@ -112,7 +113,6 @@ def analyze(text):
|
|
| 112 |
|
| 113 |
if i in prob_map:
|
| 114 |
score = prob_map[i]
|
| 115 |
-
# Standard coloring thresholds (Independent of the final label threshold)
|
| 116 |
if score < 0.35: color, bg = "#11823b", "rgba(17, 130, 59, 0.15)"
|
| 117 |
elif score < 0.70: color, bg = "#b8860b", "rgba(184, 134, 11, 0.15)"
|
| 118 |
else: color, bg = "#b80d0d", "rgba(184, 13, 13, 0.15)"
|
|
@@ -127,7 +127,7 @@ def analyze(text):
|
|
| 127 |
|
| 128 |
highlighted_html += "</div>"
|
| 129 |
|
| 130 |
-
#
|
| 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]})
|
|
|
|
| 17 |
def get_model():
|
| 18 |
global tokenizer, model
|
| 19 |
if model is None:
|
| 20 |
+
# Loading messages to help debug in HF logs
|
| 21 |
+
print(f"Loading model: {MODEL_NAME} on {device}")
|
| 22 |
tokenizer = AutoTokenizer.from_pretrained(MODEL_NAME)
|
| 23 |
|
| 24 |
+
# Use float32 on CPU to prevent build-time precision hangs
|
| 25 |
dtype = torch.float32
|
| 26 |
if device.type == "cuda" and torch.cuda.is_bf16_supported():
|
| 27 |
dtype = torch.bfloat16
|
|
|
|
| 31 |
).to(device).eval()
|
| 32 |
return tokenizer, model
|
| 33 |
|
|
|
|
| 34 |
THRESHOLD = 0.20
|
| 35 |
|
| 36 |
# -----------------------------
|
|
|
|
| 68 |
return final_blocks
|
| 69 |
|
| 70 |
# -----------------------------
|
| 71 |
+
# ANALYSIS
|
| 72 |
# -----------------------------
|
| 73 |
@torch.inference_mode()
|
| 74 |
def analyze(text):
|
|
|
|
| 113 |
|
| 114 |
if i in prob_map:
|
| 115 |
score = prob_map[i]
|
|
|
|
| 116 |
if score < 0.35: color, bg = "#11823b", "rgba(17, 130, 59, 0.15)"
|
| 117 |
elif score < 0.70: color, bg = "#b8860b", "rgba(184, 134, 11, 0.15)"
|
| 118 |
else: color, bg = "#b80d0d", "rgba(184, 13, 13, 0.15)"
|
|
|
|
| 127 |
|
| 128 |
highlighted_html += "</div>"
|
| 129 |
|
| 130 |
+
# Updated Label Logic per your request
|
| 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]})
|