File size: 3,603 Bytes
1be017e |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 |
import spaces
import gradio as gr
from transformers import AutoModelForCausalLM, AutoTokenizer
import torch
class ModelProcessor:
def __init__(self, repo_id="HuggingFaceTB/cosmo-1b"):
self.device = "cuda:0" if torch.cuda.is_available() else "cpu"
self.tokenizer = AutoTokenizer.from_pretrained(repo_id, use_fast=True)
self.model = AutoModelForCausalLM.from_pretrained(
repo_id, torch_dtype=torch.float16, device_map={"": self.device}, trust_remote_code=True
)
self.model.eval()
self.tokenizer.pad_token = self.tokenizer.eos_token
@torch.inference_mode()
def process_data_and_compute_statistics(self, prompt):
tokens = self.tokenizer(
prompt, return_tensors="pt", truncation=True, max_length=512
).to(self.model.device)
outputs = self.model(tokens["input_ids"])
logits = outputs.logits
shifted_labels = tokens["input_ids"][..., 1:].contiguous()
shifted_logits = logits[..., :-1, :].contiguous()
shifted_probs = torch.softmax(shifted_logits, dim=-1)
shifted_log_probs = torch.log_softmax(shifted_logits, dim=-1)
entropy = -torch.sum(shifted_probs * shifted_log_probs, dim=-1).squeeze()
logits_flat = shifted_logits.view(-1, shifted_logits.size(-1))
labels_flat = shifted_labels.view(-1)
probabilities_flat = torch.softmax(logits_flat, dim=-1)
true_class_probabilities = probabilities_flat.gather(
1, labels_flat.unsqueeze(1)
).squeeze(1)
nll = -torch.log(
true_class_probabilities.clamp(min=1e-9)
)
ranks = (
shifted_logits.argsort(dim=-1, descending=True)
== shifted_labels.unsqueeze(-1)
).nonzero()[:, -1]
if entropy.clamp(max=4).median() < 2.0:
return 1
return 1 if (ranks.clamp(max=4) * nll.clamp(max=4)).mean() < 5.2 else 0
processor = ModelProcessor()
@spaces.GPU(duration=180)
def detect(prompt):
prediction = processor.process_data_and_compute_statistics(prompt)
if prediction == 1:
return "<div class='output-text'>The text is likely <b>generated</b> by a language model.</div>"
else:
return "<div class='output-text'>The text is likely <b>not generated</b> by a language model.</div>"
with gr.Blocks(
css="""
.gradio-container {
max-width: 800px;
margin: 0 auto;
}
.gr-box {
box-shadow: 0 2px 4px rgba(0, 0, 0, 0.1);
padding: 20px;
border-radius: 4px;
}
.gr-button {
background-color: #007bff;
color: white;
padding: 10px 20px;
border-radius: 4px;
}
.gr-button:hover {
background-color: #0056b3;
}
.hyperlinks a {
margin-right: 10px;
}
.output-text {
text-align: center;
font-size: 24px;
font-weight: bold;
}
"""
) as demo:
with gr.Row():
with gr.Column(scale=3):
gr.Markdown("# ENTELL Model Detection - ChatGPTBots.net")
with gr.Column(scale=1):
gr.HTML(
"""
""",
elem_classes="hyperlinks",
)
with gr.Row():
with gr.Column():
prompt = gr.Textbox(
lines=8,
placeholder="Type your prompt here...",
label="Prompt",
)
submit_btn = gr.Button("Submit", variant="primary")
output = gr.HTML() # Changed to gr.HTML() to support custom HTML
submit_btn.click(fn=detect, inputs=promp |