File size: 3,295 Bytes
31c6aa2
e2d8b9b
 
 
31c6aa2
 
e2d8b9b
2735c1a
 
 
 
 
e2d8b9b
 
2735c1a
 
 
 
 
e2d8b9b
 
 
 
 
676c241
e2d8b9b
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2735c1a
 
e2d8b9b
 
 
 
 
 
0c0c1de
e2d8b9b
 
 
 
 
 
 
 
 
5310029
e2d8b9b
 
 
a1f7a6b
e2d8b9b
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
"""
Gradio Space for Human-AI Text Attribution (HATA) Model
Detects whether text is human-written or AI-generated
Supports multiple African languages
"""

# --- Deterministic suppression of Gradio audio stack under Python 3.13 ---
import os
import sys
import types

os.environ["GRADIO_DISABLE_PYDUB"] = "1"

# Provide stubs so that pydub cannot fail on audioop / pyaudioop
if "audioop" not in sys.modules:
    sys.modules["audioop"] = types.ModuleType("audioop")
if "pyaudioop" not in sys.modules:
    sys.modules["pyaudioop"] = types.ModuleType("pyaudioop")

# Now it is safe to import Gradio and the rest of the stack
import gradio as gr
import torch
import numpy as np
from transformers import AutoTokenizer, AutoModelForSequenceClassification

# ----------------------------------------------------------------------
# Model configuration
# ----------------------------------------------------------------------
MODEL_NAME = "distilbert-base-multilingual-cased"  # replace with your fine-tuned HATA checkpoint if available
DEVICE = "cuda" if torch.cuda.is_available() else "cpu"

tokenizer = AutoTokenizer.from_pretrained(MODEL_NAME)
model = AutoModelForSequenceClassification.from_pretrained(MODEL_NAME, num_labels=2)
model.to(DEVICE)
model.eval()

LABELS = ["Human-written", "AI-generated"]

# ----------------------------------------------------------------------
# Inference routine
# ----------------------------------------------------------------------
@torch.no_grad()
def hata_predict(text: str):
    if not text or not text.strip():
        return {"Human-written": 0.0, "AI-generated": 0.0}

    inputs = tokenizer(
        text,
        return_tensors="pt",
        truncation=True,
        padding=True,
        max_length=512,
    ).to(DEVICE)

    outputs = model(**inputs)
    logits = outputs.logits.squeeze(0)
    probs = torch.softmax(logits, dim=-1).cpu().numpy()

    return {LABELS[i]: float(probs[i]) for i in range(len(LABELS))}

# ----------------------------------------------------------------------
# Gradio interface
# ----------------------------------------------------------------------
with gr.Blocks(title="Multilingual HATA System") as demo:
    gr.Markdown(
        """
        # Multilingual Human–AI Text Attribution (HATA)

        This system estimates whether an input passage is **human-written** or
        **AI-generated**, with a focus on multilingual and African-language use
        cases (e.g., Hausa, Yoruba, Igbo, Pidgin).

        The backend is a Transformer-based classifier fine-tuned for attribution.
        """
    )

    with gr.Row():
        with gr.Column(scale=3):
            text_input = gr.Textbox(
                label="Input Text",
                placeholder="Paste a paragraph in Hausa, Yoruba, Igbo, Pidgin, or English...",
                lines=8,
            )
            submit_btn = gr.Button("Analyze")
        with gr.Column(scale=2):
            output = gr.Label(label="Attribution Probabilities")

    submit_btn.click(
        fn=hata_predict,
        inputs=text_input,
        outputs=output,
    )

# ----------------------------------------------------------------------
# Entry point
# ----------------------------------------------------------------------
if __name__ == "__main__":
    demo.launch()