app.py

import gradio as gr
from transformers import AutoTokenizer, AutoModelForCausalLM
import lftk
import spacy
import time
import os
import openai

# Load the Vicuna 7B model and tokenizer
vicuna_tokenizer = AutoTokenizer.from_pretrained("lmsys/vicuna-7b-v1.3")
vicuna_model = AutoModelForCausalLM.from_pretrained("lmsys/vicuna-7b-v1.3")

# Load the LLaMA 7b model and tokenizer
llama_tokenizer = AutoTokenizer.from_pretrained("daryl149/llama-2-7b-chat-hf")
llama_model = AutoModelForCausalLM.from_pretrained("daryl149/llama-2-7b-chat-hf")

template_single = '''Please output any <{}> in the following sentence one per line without any additional text: "{}"'''

# def linguistic_features(message):
#     # Load a trained spaCy pipeline
#     nlp = spacy.load("en_core_web_sm")

#     # Create a spaCy doc object
#     doc = nlp(message)

#     # Initiate LFTK extractor by passing in the doc
#     LFTK_extractor = lftk.Extractor(docs=doc)

#     # Customize LFTK extractor (optional)
#     LFTK_extractor.customize(stop_words=True, punctuations=False, round_decimal=3)

#     # Use LFTK to dynamically extract handcrafted linguistic features
#     features_to_extract = lftk.search_features(family="wordsent", language="general", return_format="list_key")
#     extracted_features = LFTK_extractor.extract(features=features_to_extract)

#     print('Linguistic Features:', extracted_features)

#     return extracted_features


def update_api_key(new_key):
    global api_key
    os.environ['OPENAI_API_TOKEN'] = new_key
    openai.api_key = os.environ['OPENAI_API_TOKEN']

def chat(system_prompt, user_prompt, model = 'gpt-3.5-turbo', temperature = 0, verbose = False):
    ''' Normal call of OpenAI API '''
    response = openai.ChatCompletion.create(
    temperature = temperature,
    model=model,
    messages=[
        {"role": "system", "content": system_prompt},
        {"role": "user", "content": user_prompt}
    ])
    
    res = response['choices'][0]['message']['content']
    
    if verbose:
        print('System prompt:', system_prompt)
        print('User prompt:', user_prompt)
        print('GPT response:', res)
        
    return res

def format_chat_prompt(message, chat_history, max_convo_length):
    prompt = ""
    for turn in chat_history[-max_convo_length:]:
        user_message, bot_message = turn
        prompt = f"{prompt}\nUser: {user_message}\nAssistant: {bot_message}"
    prompt = f"{prompt}\nUser: {message}\nAssistant:"
    return prompt

def gpt_respond(tab_name, message, chat_history, max_convo_length = 10):
    formatted_prompt = format_chat_prompt(message, chat_history, max_convo_length)
    print('Prompt + Context:')
    print(formatted_prompt)
    bot_message = chat(system_prompt = f'''Generate the output only for the assistant. Please output any <{tab_name}> in the following sentence one per line without any additional text.''',
                       user_prompt = formatted_prompt)
    chat_history.append((message, bot_message))
    return "", chat_history

def vicuna_respond(tab_name, message, chat_history):
    formatted_prompt = f'''Generate the output only for the assistant. Please output any {tab_name} in the following sentence one per line without any additional text: {message}'''
    print('Vicuna Ling Ents Fn - Prompt + Context:')
    print(formatted_prompt)
    input_ids = vicuna_tokenizer.encode(formatted_prompt, return_tensors="pt")
    output_ids = vicuna_model.generate(input_ids, do_sample=True, max_length=1024, num_beams=5, no_repeat_ngram_size=2)
    bot_message = vicuna_tokenizer.decode(output_ids[0], skip_special_tokens=True)
    print(bot_message)
    
    # Remove formatted prompt from bot_message
    bot_message = bot_message.replace(formatted_prompt, '')
    print(bot_message)
        
    chat_history.append((formatted_prompt, bot_message))
    time.sleep(2)
    return tab_name, "", chat_history

def llama_respond(tab_name, message, chat_history):
    formatted_prompt = f'''Generate the output only for the assistant. Please output any {tab_name} in the following sentence one per line without any additional text: {message}'''
    # print('Llama - Prompt + Context:')
    # print(formatted_prompt)
    input_ids = llama_tokenizer.encode(formatted_prompt, return_tensors="pt")
    output_ids = llama_model.generate(input_ids, do_sample=True, max_length=1024, num_beams=5, no_repeat_ngram_size=2)
    bot_message = llama_tokenizer.decode(output_ids[0], skip_special_tokens=True)

    # Remove formatted prompt from bot_message
    bot_message = bot_message.replace(formatted_prompt, '')
    # print(bot_message)
        
    chat_history.append((formatted_prompt, bot_message))
    time.sleep(2)
    return tab_name, "", chat_history

def gpt_strategies_respond(strategy, task_name, task_ling_ent, message, chat_history, max_convo_length = 10):
    formatted_system_prompt = ""
    if (task_name == "POS Tagging"):
        if (strategy == "S1"):
            formatted_system_prompt = f'''Generate the output only for the assistant. Please output any {task_ling_ent} in the following sentence one per line without any additional text: {message}'''
        elif (strategy == "S2"):
            formatted_system_prompt = f'''Please POS tag the following sentence using Universal POS tag set without generating any additional text: {message}'''
        elif (strategy == "S3"):
            formatted_system_prompt = f'''Please POS tag the following sentence using Universal POS tag set without generating any additional text: {message}'''
    elif (task_name == "Chunking"):
        if (strategy == "S1"):
            formatted_system_prompt = f'''Generate the output only for the assistant. Please output any {task_ling_ent} in the following sentence one per line without any additional text: {message}'''
        elif (strategy == "S2"):
            formatted_system_prompt = f'''Please chunk the following sentence in CoNLL 2000 format with BIO tags without outputing any additional text: {message}'''
        elif (strategy == "S3"):
            formatted_system_prompt = f'''Please chunk the following sentence in CoNLL 2000 format with BIO tags without outputing any additional text: {message}'''
    
    formatted_prompt = format_chat_prompt(message, chat_history, max_convo_length)
    print('Prompt + Context:')
    print(formatted_prompt)
    bot_message = chat(system_prompt = formatted_system_prompt,
                       user_prompt = formatted_prompt)
    chat_history.append((message, bot_message))
    return "", chat_history

def vicuna_strategies_respond(strategy, task_name, task_ling_ent, message, chat_history):
    formatted_prompt = ""
    if (task_name == "POS Tagging"):
        if (strategy == "S1"):
            formatted_prompt = f'''Generate the output only for the assistant. Please output any {task_ling_ent} in the following sentence one per line without any additional text: {message}'''
        elif (strategy == "S2"):
            formatted_prompt = f'''Please POS tag the following sentence using Universal POS tag set without generating any additional text: {message}'''
        elif (strategy == "S3"):
            formatted_prompt = f'''Please POS tag the following sentence using Universal POS tag set without generating any additional text: {message}'''
    elif (task_name == "Chunking"):
        if (strategy == "S1"):
            formatted_prompt = f'''Generate the output only for the assistant. Please output any {task_ling_ent} in the following sentence one per line without any additional text: {message}'''
        elif (strategy == "S2"):
            formatted_prompt = f'''Please chunk the following sentence in CoNLL 2000 format with BIO tags without outputing any additional text: {message}'''
        elif (strategy == "S3"):
            formatted_prompt = f'''Please chunk the following sentence in CoNLL 2000 format with BIO tags without outputing any additional text: {message}'''
    
    print('Vicuna Strategy Fn - Prompt + Context:')
    print(formatted_prompt)
    input_ids = vicuna_tokenizer.encode(formatted_prompt, return_tensors="pt")
    output_ids = vicuna_model.generate(input_ids, do_sample=True, max_length=1024, num_beams=5, no_repeat_ngram_size=2)
    bot_message = vicuna_tokenizer.decode(output_ids[0], skip_special_tokens=True)
    print(bot_message)
    
    # Remove formatted prompt from bot_message
    bot_message = bot_message.replace(formatted_prompt, '')
    print(bot_message)
        
    chat_history.append((formatted_prompt, bot_message))
    time.sleep(2)
    return task_name, "", chat_history

def llama_strategies_respond(strategy, task_name, task_ling_ent, message, chat_history):
    formatted_prompt = ""
    if (task_name == "POS Tagging"):
        if (strategy == "S1"):
            formatted_prompt = f'''Generate the output only for the assistant. Please output any {task_ling_ent} in the following sentence one per line without any additional text: {message}'''
        elif (strategy == "S2"):
            formatted_prompt = f'''Please POS tag the following sentence using Universal POS tag set without generating any additional text: {message}'''
        elif (strategy == "S3"):
            formatted_prompt = f'''Please POS tag the following sentence using Universal POS tag set without generating any additional text: {message}'''
    elif (task_name == "Chunking"):
        if (strategy == "S1"):
            formatted_prompt = f'''Generate the output only for the assistant. Please output any {task_ling_ent} in the following sentence one per line without any additional text: {message}'''
        elif (strategy == "S2"):
            formatted_prompt = f'''Please chunk the following sentence in CoNLL 2000 format with BIO tags without outputing any additional text: {message}'''
        elif (strategy == "S3"):
            formatted_prompt = f'''Please chunk the following sentence in CoNLL 2000 format with BIO tags without outputing any additional text: {message}'''
    
    # print('Llama Strategies - Prompt + Context:')
    # print(formatted_prompt)
    input_ids = llama_tokenizer.encode(formatted_prompt, return_tensors="pt")
    output_ids = llama_model.generate(input_ids, do_sample=True, max_length=1024, num_beams=5, no_repeat_ngram_size=2)
    bot_message = llama_tokenizer.decode(output_ids[0], skip_special_tokens=True)
    # print(bot_message)
    
    # Remove formatted prompt from bot_message
    bot_message = bot_message.replace(formatted_prompt, '')
    # print(bot_message)
        
    chat_history.append((formatted_prompt, bot_message))
    time.sleep(2)
    return task_name, "", chat_history

def interface():

        # prompt = template_single.format(tab_name, textbox_prompt)

        with gr.Tab("Linguistic Entities"):
            gr.Markdown(" Description Here ")

            # Inputs
            ling_ents_prompt = gr.Textbox(show_label=False, placeholder="Write a prompt and press enter")
            with gr.Row():
                # Will activate after getting API key
                ling_ents_apikey_input = gr.Textbox(label="Open AI Key", placeholder="Enter your Openai key here", type="password")
                ling_ents_apikey_btn = gr.Button(value="Submit Key", scale=0)
            linguistic_entities = gr.Dropdown(["Noun", "Determiner", "Noun phrase", "Verb phrase", "Dependent clause", "T-units"], label="Linguistic Entity")
            ling_ents_btn = gr.Button(value="Submit")

            # Outputs
            gr.Markdown("Strategy 1 QA-Based Prompting")

            linguistic_features_textbox = gr.Textbox(label="Linguistic Features", disabled=True)

            with gr.Row():
                vicuna_ling_ents_chatbot = gr.Chatbot(label="vicuna-7b")
                llama_ling_ents_chatbot = gr.Chatbot(label="llama-7b")
                gpt_ling_ents_chatbot = gr.Chatbot(label="gpt-3.5")
            clear = gr.ClearButton(components=[ling_ents_prompt, ling_ents_apikey_input, vicuna_ling_ents_chatbot, llama_ling_ents_chatbot, gpt_ling_ents_chatbot])

            # Event Handler for Vicuna Chatbot
            ling_ents_btn.click(vicuna_respond, inputs=[linguistic_entities, ling_ents_prompt, vicuna_ling_ents_chatbot], 
                      outputs=[linguistic_entities, ling_ents_prompt, vicuna_ling_ents_chatbot])
            
            # Event Handler for LLaMA Chatbot
            ling_ents_btn.click(llama_respond, inputs=[linguistic_entities, ling_ents_prompt, llama_ling_ents_chatbot], 
                      outputs=[linguistic_entities, ling_ents_prompt, llama_ling_ents_chatbot])
            
            # Event Handler for GPT 3.5 Chatbot, user must submit api key before submitting the prompt
            # Will activate after getting API key
            # ling_ents_apikey_btn.click(update_api_key, inputs=ling_ents_apikey_input)
            # ling_ents_btn.click(gpt_respond, inputs=[linguistic_entities, ling_ents_prompt, gpt_ling_ents_chatbot], 
            #                     outputs=[linguistic_entities, ling_ents_prompt, gpt_ling_ents_chatbot])

        with gr.Tab("POS/Chunking"):
            gr.Markdown(" Description Here ")

            # Inputs
            task_prompt = gr.Textbox(show_label=False, placeholder="Write a prompt and press enter")
            with gr.Row():
                task_apikey_input = gr.Textbox(label="Open AI Key", placeholder="Enter your Openai key here", type="password")
                task_apikey_btn = gr.Button(value="Submit Key", scale=0)
            task = gr.Dropdown(["POS Tagging", "Chunking"], label="Task")
            task_linguistic_entities = gr.Dropdown(["Noun", "Determiner", "Noun phrase", "Verb phrase", "Dependent clause", "T-units"], label="Linguistic Entity For Strategy 1")
            task_btn = gr.Button(value="Submit")

            # Outputs
            gr.Markdown("Strategy 1 QA-Based Prompting")
            strategy1 = gr.Markdown("S1", visible=False)
            with gr.Row():
                vicuna_S1_chatbot = gr.Chatbot(label="vicuna-7b")
                llama_S1_chatbot = gr.Chatbot(label="llama-7b")
                gpt_S1_chatbot = gr.Chatbot(label="gpt-3.5")
            gr.Markdown("Strategy 2 Instruction-Based Prompting")
            strategy2 = gr.Markdown("S2", visible=False)
            with gr.Row():
                vicuna_S2_chatbot = gr.Chatbot(label="vicuna-7b")
                llama_S2_chatbot = gr.Chatbot(label="llama-7b")
                gpt_S2_chatbot = gr.Chatbot(label="gpt-3.5")
            gr.Markdown("Strategy 3 Structured Prompting")
            strategy3 = gr.Markdown("S3", visible=False)
            with gr.Row():
                vicuna_S3_chatbot = gr.Chatbot(label="vicuna-7b")
                llama_S3_chatbot = gr.Chatbot(label="llama-7b")
                gpt_S3_chatbot = gr.Chatbot(label="gpt-3.5")
            clear_all = gr.ClearButton(components=[task_prompt, task_apikey_input,
                                               vicuna_S1_chatbot, llama_S1_chatbot, gpt_S1_chatbot, 
                                               vicuna_S2_chatbot, llama_S2_chatbot, gpt_S2_chatbot,
                                               vicuna_S3_chatbot, llama_S3_chatbot, gpt_S3_chatbot])
            
            # vicuna_strategies_respond(strategy, task_name, task_ling_ent, message, chat_history):
            # Event Handlers for Vicuna Chatbot POS/Chunk
            task_btn.click(vicuna_strategies_respond, inputs=[strategy1, task, task_linguistic_entities, task_prompt, vicuna_S1_chatbot], 
                      outputs=[task, task_prompt, vicuna_S1_chatbot])
            task_btn.click(vicuna_strategies_respond, inputs=[strategy2, task, task_linguistic_entities, task_prompt, vicuna_S2_chatbot], 
                      outputs=[task, task_prompt, vicuna_S2_chatbot])
            task_btn.click(vicuna_strategies_respond, inputs=[strategy3, task, task_linguistic_entities, task_prompt, vicuna_S3_chatbot], 
                      outputs=[task, task_prompt, vicuna_S3_chatbot])
            
            # Event Handler for LLaMA Chatbot POS/Chunk
            task_btn.click(llama_strategies_respond, inputs=[strategy1, task, task_linguistic_entities, task_prompt, llama_S1_chatbot], 
                      outputs=[task, task_prompt, llama_S1_chatbot])
            task_btn.click(llama_strategies_respond, inputs=[strategy2, task, task_linguistic_entities, task_prompt, llama_S2_chatbot], 
                      outputs=[task, task_prompt, llama_S2_chatbot])
            task_btn.click(llama_strategies_respond, inputs=[strategy3, task, task_linguistic_entities, task_prompt, llama_S3_chatbot], 
                      outputs=[task, task_prompt, llama_S3_chatbot])
            
            # Event Handler for GPT 3.5 Chatbot POS/Chunk, user must submit api key before submitting the prompt
            # Will activate after getting API key
            # task_apikey_btn.click(update_api_key, inputs=ling_ents_apikey_input)
            # task_btn.click(gpt_strategies_respond, inputs=[strategy1, task, task_linguistic_entities, gpt_S1_chatbot], 
            #                     outputs=[task, task_prompt, gpt_S1_chatbot])
            # task_btn.click(gpt_strategies_respond, inputs=[strategy1, task, task_linguistic_entities, gpt_S2_chatbot], 
            #                     outputs=[task, task_prompt, gpt_S2_chatbot])
            # task_btn.click(gpt_strategies_respond, inputs=[strategy1, task, task_linguistic_entities, gpt_S3_chatbot], 
            #                     outputs=[task, task_prompt, gpt_S3_chatbot])
            

        # textbox_prompt.submit(vicuna_respond, inputs=[textbox_prompt, vicuna_S1_chatbot], outputs=[textbox_prompt, vicuna_S1_chatbot])
        # textbox_prompt.submit(respond, inputs=[textbox_prompt, vicuna_S2_chatbot], outputs=[textbox_prompt, vicuna_S2_chatbot])
        # textbox_prompt.submit(respond, inputs=[textbox_prompt, vicuna_S3_chatbot], outputs=[textbox_prompt, vicuna_S3_chatbot])

        #textbox_prompt.submit(llama_respond, inputs=[textbox_prompt, llama_S1_chatbot], outputs=[textbox_prompt, llama_S1_chatbot])

        # btn.click(lambda _, 
        #           message=textbox_prompt: linguistic_features_textbox.update(linguistic_features(textbox_prompt.value)),
        #           inputs=[textbox_prompt], 
        #           outputs=[linguistic_features_textbox])

with gr.Blocks() as demo:
    gr.Markdown("# LLM Evaluator With Linguistic Scrutiny")

    # load interface
    interface()
    
demo.launch()