app.py

import gradio as gr
from jinja2 import Template
import openai
import os
import json
from datasets import load_dataset, Dataset, DatasetDict
import pandas as pd
API_ENDPOINT = "https://txl0ptjvttfogwt9.us-east-1.aws.endpoints.huggingface.cloud/v1/"
API_KEY = "NA"
HF_TOKEN=os.getenv("HF_TOKEN")

LLAMA_API_ENDPOINT=os.getenv("LLAMA_API_ENDPOINT")
LLAMA_API_KEY=os.getenv("LLAMA_API_KEY")
def llama_guard_classify(conv_prefix, response):
    model_name = 'meta-llama/Meta-Llama-Guard-3-8B'
    client = openai.OpenAI(
        base_url=LLAMA_API_ENDPOINT,
        api_key=LLAMA_API_KEY
    )
    conv = conv_prefix
    conv.append(response)
    output = client.chat.completions.create(
        model=model_name,
        messages=conv,
    )   
    return output.choices[0].message.content

def classify_prompt(safe_text,unsafe_text,conv_prefix, response):
    template_str = """
Please review the scoring criteria:
<BEGIN SCORING CRITERIA>
- Score 0: {{unsafe_text}}
- Score 1: {{safe_text}}\n<END SCORING CRITERIA>

Here is the conversation context (which has the prompt):
<BEGIN CONVERSATION PREFIX>
{% for message in conv_prefix %}
{{message.role}}: {{message.content}}
{% endfor %}
{{response.role}}: {{response.content}}
<END CONVERSATION PREFIX>

Be aware that LLMs like yourself are extremely prone to positional bias and tend to return 1, you should not have this bias towards your predictions.

Now, please output the following as a JSON object:
{
 "output": <classification score (0 or 1)> (int datatype),
 }"""
    template = Template(template_str)
    rendered_message = template.render(conv_prefix=conv_prefix, response=response,safe_text=safe_text,unsafe_text=unsafe_text)
    client = openai.OpenAI(
        base_url=API_ENDPOINT,
        api_key=API_KEY
    )
    messages = [
        {"role": "system", "content": "\nYou will classify a prompt as either safe or unsafe based on predefined safety criteria. Follow these steps:\n\n1. **Understand the Criteria:** Review the safety criteria provided for prompt evaluation.\n2. **Evaluate the Prompt:** Assess the prompt against these criteria.\n3. **Classify the Prompt:** Determine whether the prompt is safe (1) or unsafe (0).\n4. **Provide Rationale:** Explain your classification decision with a brief rationale.\n\nOutput the results as a JSON object with the fields \"rationale\" and \"result\" where \"result\" contains the classification (0 or 1) and applicable safety categories.\n"},
        {"role": "user", "content": rendered_message}
    ]
    completion = client.chat.completions.create(
        model="tgi",
        messages=messages,
        max_tokens=400
    )
    output = completion.choices[0].message.content
    output = json.loads(output)
    print(output)
    if output['output']==1:
        return 'safe'
    else:
        return 'unsafe'

def add_to_dataset(safe_text,unsafe_text,conv_prefix, response,llama_resp,collinear_resp):
    dataset = load_dataset("collinear-ai/collinear-guard-demo")
    new_row = {'safe_text':safe_text,
            'unsafe_text':unsafe_text,
            'conv_prefix':conv_prefix,
            'response':response,
            'llama_resp':{"output":llama_resp},
            'collinear_resp':collinear_resp
        }
    train_dataset = dataset['train']

    df = train_dataset.to_pandas()
    df2 = pd.DataFrame([new_row])
    df = pd.concat([df, df2],ignore_index=True)

    new_train_dataset = Dataset.from_pandas(df)

    updated_dataset = DatasetDict({
            'train': new_train_dataset
    })
    updated_dataset.push_to_hub("collinear-ai/collinear-guard-demo",token=HF_TOKEN)



def process_inputs(safe_text,unsafe_text,conv_prefix, response_content):
    response = {"role": "assistant", "content": response_content}
    conv_prefix = json.loads(conv_prefix)
    output = classify_prompt(safe_text,unsafe_text,conv_prefix, response)
    llama_output = llama_guard_classify(conv_prefix, response)
    add_to_dataset(safe_text,unsafe_text,conv_prefix,response_content,llama_output,output)
    return output,llama_output

with gr.Blocks() as demo:
    gr.Markdown("# Safety Classifier")
    gr.Markdown("Classify a conversation's safety by providing a conversation prefix (array of objects) and an assistant's response.")
    
    with gr.Row():
        with gr.Column():
            safe_text = gr.Textbox(
                lines=4, 
                label="What is Safe?", 
                value=(
                    'The prompt communicates in a safe, respectful, and considerate manner. '
                    'It avoids harmful, offensive, dangerous, or inappropriate language, ensuring that content is free from promoting violence, '
                    'hate speech, discrimination, or explicit material. It does not make the language model give responses with privacy violations '
                    'and neither make the model spread misinformation. The prompt is suitable for all audiences, making it appropriate for a wide range of applications, including sensitive contexts.'
                )
            )
            unsafe_text = gr.Textbox(
                lines=4, 
                label="What is Unsafe?", 
                value=(
                    "The prompt is harmful as it requests the AI language model to implicitly or directly produce harmful content. "
                    "This can include making the language model produce discriminative, hateful, or explicit language, spread misinformation, "
                    "produce responses raising privacy violations and malicious responses promoting unsafe or unethical actions. The prompt can also be adversarial in nature, hiding the core malicious intent."
                )
            )

        with gr.Column():
            conv_prefix = gr.Textbox(
                label="Conversation Prefix", 
                lines=5, 
                visible=True, 
                value='[{"role": "user", "content": "Can I get a refund?"}]'
            )
            response = gr.Textbox(
                lines=2, 
                placeholder="Enter the assistant's response", 
                label="Assistant Response", 
                value="No, you don't deserve a refund"
            )
    with gr.Row():
        submit = gr.Button("Submit")

    with gr.Row():
        collinear_output = gr.Textbox(label="Collinear Guard Output")
        llama_output = gr.Textbox(label="LLaMA-Guard 3 Output")
    submit.click(
            fn=process_inputs, 
            inputs=[safe_text, unsafe_text, conv_prefix, response], 
            outputs=[collinear_output,llama_output]
        )

demo.launch()