app.py

import gradio as gr
import joblib
import torch
import sys
import os
sys.path.append(os.path.dirname(os.path.abspath(__file__)))
from l3prune import LLMEncoder
#load the model
best_clf = joblib.load("./saved/classifier_llama32.joblib")


encoder = LLMEncoder.from_pretrained(
    "./saved/pruned_encoder_llama32",
    device_map="cpu",
    torch_dtype=torch.bfloat16,
    #torch_dtype=torch,
    #cache_dir=cache_dir
)



def classify_prompt(prompt):
    #response = client.text_classification(prompt)
    #label = response[0]['label']
    #score = response[0]['score']
    
    #if label == 'hate':
    #    result = f"Harmful (Confidence: {score:.2%})"
    #else:
    #    result = f"Benign (Confidence: {score:.2%})"

    X = encoder.encode([prompt])
    result = best_clf.predict(X)[0]
    
    return "Harmful" if result else "Benign"

demo = gr.Interface(
    fn=classify_prompt,
    inputs=gr.Textbox(lines=3, placeholder="Enter a prompt to classify..."),
    outputs=gr.Textbox(label="Classification Result"),
    title="Harmful Prompt Classifier",
    description="This app classifies whether a given prompt is potentially harmful or benign.",
    show_api=False,
    show_response_timing=True
)
if __name__ == "__main__":
    demo.launch()