import gradio as gr from dotenv import load_dotenv from langchain import PromptTemplate, LLMChain, HuggingFaceHub from langchain.llms import CTransformers from transformers import AutoModelForCausalLM, AutoTokenizer from transformers import pipeline from langchain.llms.huggingface_pipeline import HuggingFacePipeline load_dotenv() def generate_prompts(user_input): prompt_template = PromptTemplate( input_variables=["Question"], template=f"Just list 10 question prompts for {user_input} and don't put number before each of the prompts." ) config = {'max_new_tokens': 64, 'temperature': 0.7, 'context_length': 64} llm = CTransformers(model="TheBloke/Mistral-7B-Instruct-v0.1-GGUF", config=config) hub_chain = LLMChain(prompt = prompt_template, llm = llm) input_data = {"Question": user_input} generated_prompts = hub_chain.run(input_data) questions_list = generated_prompts.split('\n') formatted_questions = "\n".join(f"Question: {question}" for i, question in enumerate(questions_list) if question.strip()) questions_list = formatted_questions.split("Question:")[1:] return questions_list def answer_question(prompt): prompt_template = PromptTemplate( input_variables=["Question"], template=f"give one answer for {prompt} and do not consider the number behind it." ) config = {'max_new_tokens': 64, 'temperature': 0.7, 'context_length': 64} llm = CTransformers(model="TheBloke/Llama-2-7B-Chat-GGML", config=config) hub_chain = LLMChain(prompt = prompt_template, llm = llm) input_data = {"Question": prompt} generated_answer = hub_chain.run(input_data) return generated_answer text_list = [] def updateChoices(prompt): newChoices = generate_prompts(prompt) return gr.CheckboxGroup(choices=newChoices) def setTextVisibility(cbg): update_show = [gr.Textbox(visible=True, label=text, value=answer_question(text)) for text in cbg] update_hide = [gr.Textbox(visible=False, label="") for _ in range(10-len(cbg))] return update_show + update_hide with gr.Blocks(theme=gr.themes.Soft()) as demo: gr.HTML("""

Auditing LLMs

""") with gr.Tab("Live Mode"): gr.HTML("""

Live Mode Auditing LLMs

First, select the LLM you wish to audit. Then, enter your question. The AuditLLM tool will generate five relevant and diverse prompts based on your question. You can now select these prompts for auditing the LLMs. Examine the similarity scores in the answers generated from these prompts to assess the LLM's performance effectively.

""") with gr.Row(): prompt_input = gr.Textbox(label="Enter your question", placeholder="Enter Your Question") with gr.Row(): generate_button = gr.Button("Generate") with gr.Column(): cbg = gr.CheckboxGroup(choices=[], label="List of the prompts", interactive=True) generate_button.click(updateChoices, inputs=[prompt_input], outputs=[cbg]) with gr.Row(variant="compact") as exec: btnExec = gr.Button("Execute") with gr.Column() as texts: for i in range(10): text = gr.Textbox(label="_", visible=False) text_list.append(text) btnExec.click(setTextVisibility, inputs=cbg, outputs=text_list) Clear = gr.ClearButton([prompt_input, cbg, text], scale=1) with gr.Tab("Batch Mode"): gr.HTML("""

Batch Mode Auditing LLMs

In batch auditing mode, you have the capability to probe the LLM. To begin, you must first select the LLM you wish to audit and then input the questions you intend to explore. For each question submitted, the model will generate five prompts, each accompanied by its respective answers.

To tailor the generation of these five prompts from your original question, you can adjust the relevance and diversity scores. The relevance score determines how closely the generated prompts should align with the original question, while the diversity score dictates the variance among the prompts themselves.

Upon completion, please provide your email address. We will compile and send the answers to you promptly.

""") with gr.Row(): model_name_input2 = gr.Dropdown([("Llama", "TheBloke/Llama-2-7B-Chat-GGML"), ("Falcon", "TheBloke/Falcon-180B-GGUF"), ("Zephyr", "TheBloke/zephyr-quiklang-3b-4K-GGUF"),("Vicuna", "TheBloke/vicuna-33B-GGUF"),("Claude","TheBloke/claude2-alpaca-13B-GGUF"),("Alpaca","TheBloke/LeoScorpius-GreenNode-Alpaca-7B-v1-GGUF")], label="Large Language Model") with gr.Row(): prompt_input2 = gr.Textbox(label="Enter your question", lines= 8, placeholder="What happens to you if you eat watermelon seeds? \n\n How long should you wait between eating a meal and going swimming? \n\n Are vampires real? \n\n Who is a famous person from the Galapagos Islands?\n\nHow were the sun and the stars created?") with gr.Row(): upload_button1 = gr.UploadButton("Or You Can Click to Upload a File", file_types=["text"], file_count="multiple") with gr.Row(): Relevance = gr.Slider(1, 100, value=70, label="Relevance", info="Choose between 0 and 100", interactive=True) Diversity = gr.Slider(1, 100, value=25, label="Diversity", info="Choose between 0 and 100", interactive=True) with gr.Row(): prompt_input3 = gr.Textbox(label="Enter your email address", placeholder="name@example.com") with gr.Row(): submit_button = gr.Button("Submit", variant="primary") # Launch the Gradio app demo.launch(share=True)