app.py

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

# 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")

def linguistic_features_fn(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 = lftk.Extractor(docs=doc)

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

    # Use LFTK to dynamically extract handcrafted linguistic features
    extracted_features = LFTK.extract(features = ["a_word_ps", "a_kup_pw", "n_noun"])

    formatted_output = json.dumps(extracted_features, indent=2)

    print(formatted_output)

    return formatted_output

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

def chat(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": "user", "content": user_prompt}
    ])
    
    res = response['choices'][0]['message']['content']
    
    if verbose:
        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(have_key, tab_name, message, chat_history, max_convo_length = 10):
    if (have_key == "No"):
        return "", chat_history
    
    formatted_prompt = format_chat_prompt(message, chat_history, max_convo_length)
    print('GPT ling ents Prompt + Context:')
    print(formatted_prompt)
    bot_message = chat(user_prompt = f'''Output any <{tab_name}> in the following sentence: "{formatted_prompt}"''')
    chat_history.insert(0, (message, bot_message))
    return "", chat_history

def vicuna_respond(tab_name, message, chat_history):
    formatted_prompt = f'''Output any {tab_name} in the following sentence: "{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.insert(0, (formatted_prompt, bot_message))
    time.sleep(2)
    return tab_name, "", chat_history

def llama_respond(tab_name, message, chat_history):
    formatted_prompt = f'''Output any {tab_name} in the following sentence: "{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.insert(0, (formatted_prompt, bot_message))
    time.sleep(2)
    return tab_name, "", chat_history

def gpt_strategies_respond(have_key, strategy, task_name, task_ling_ent, message, chat_history, max_convo_length = 10):
    if (have_key == "No"):
        return "", chat_history
    
    formatted_system_prompt = ""
    if (task_name == "POS Tagging"):
        if (strategy == "S1"):
            formatted_system_prompt = f'''Output any {task_ling_ent} in the following sentence: "{message}"'''
        elif (strategy == "S2"):
            formatted_system_prompt = f'''POS tag the following sentence using Universal POS tag set: "{message}"'''
        elif (strategy == "S3"):
            with open('demonstration_3_42_pos.txt', 'r') as f:
                demon_pos = f.read()
            formatted_system_prompt = f'''"{demon_pos}". Using the POS tag structure above, POS tag the following sentence: "{message}"'''
    elif (task_name == "Chunking"):
        if (strategy == "S1"):
            formatted_system_prompt = f'''Output any {task_ling_ent} in the following sentence: "{message}"'''
        elif (strategy == "S2"):
            formatted_system_prompt = f'''Chunk the following sentence in CoNLL 2000 format with BIO tags: "{message}"'''
        elif (strategy == "S3"):
            with open('demonstration_3_42_chunk.txt', 'r') as f:
                demon_chunk = f.read()
            formatted_system_prompt = f'''"{demon_chunk}". Using the POS tag structure above, POS tag the following sentence: "{message}"'''
        
    formatted_prompt = format_chat_prompt(message, chat_history, max_convo_length)
    print('GPT coreNLP Prompt + Context:')
    print(formatted_prompt)
    bot_message = chat(user_prompt = formatted_system_prompt)
    chat_history.insert(0, (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'''Output any {task_ling_ent} in the following sentence: "{message}"'''
        elif (strategy == "S2"):
            formatted_prompt = f'''POS tag the following sentence using Universal POS tag set: "{message}"'''
        elif (strategy == "S3"):
            with open('demonstration_3_42_pos.txt', 'r') as f:
                demon_pos = f.read()
            formatted_prompt = f'''"{demon_pos}". Using the POS tag structure above, POS tag the following sentence: "{message}"'''
    elif (task_name == "Chunking"):
        if (strategy == "S1"):
            formatted_prompt = f'''Output any {task_ling_ent} in the following sentence: "{message}"'''
        elif (strategy == "S2"):
            formatted_prompt = f'''Chunk the following sentence in CoNLL 2000 format with BIO tags: "{message}"'''
        elif (strategy == "S3"):
            with open('demonstration_3_42_chunk.txt', 'r') as f:
                demon_chunk = f.read()
            formatted_prompt = f'''"{demon_chunk}". Using the Chunking structure above, Chunk the following sentence: "{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.insert(0, (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'''Output any {task_ling_ent} in the following sentence: "{message}"'''
        elif (strategy == "S2"):
            formatted_prompt = f'''POS tag the following sentence using Universal POS tag set: "{message}"'''
        elif (strategy == "S3"):
            with open('demonstration_3_42_pos.txt', 'r') as f:
                demon_pos = f.read()
            formatted_prompt = f'''"{demon_pos}". Using the POS tag structure above, POS tag the following sentence: "{message}"'''
    elif (task_name == "Chunking"):
        if (strategy == "S1"):
            formatted_prompt = f'''Output any {task_ling_ent} in the following sentence: "{message}"'''
        elif (strategy == "S2"):
            formatted_prompt = f'''Chunk the following sentence in CoNLL 2000 format with BIO tags: "{message}"'''
        elif (strategy == "S3"):
            with open('demonstration_3_42_chunk.txt', 'r') as f:
                demon_chunk = f.read()
            formatted_prompt = f'''"{demon_chunk}". Using the Chunking structure above, Chunk the following sentence: "{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.insert(0, (formatted_prompt, bot_message))
    time.sleep(2)
    return task_name, "", chat_history

def interface():
        with gr.Tab("Linguistic Entities"):
            with gr.Row():
                gr.Markdown("""
                            ## 📜 Step-By-Step Instructions

                            - Enter a sentence for three models to process (Vicuna-7b, LLaMA-7b and GPT-3.5).
                            - If you own an OpenAI API key, select 'Yes' in the dropdown. If you don't own one, select 'No'.
                                - If you selected 'Yes', enter your OpenAI API Key [Link to your OpenAI keys](https://platform.openai.com/api-keys).
                                - If you selected 'No', leave the 'OpenAI Key' field blank and continue with the rest.
                            - Select a Linguistic Entity from the Dropdown or enter a custom one.
                            - Click 'Submit' to send your inputs to the models.
                            - To submit a new prompt, repeat all the steps above and click 'Submit' again. Your new prompt should appear on the top of previous ones.

                            ### ⏳ After you click 'Submit', the models will take a couple seconds to process your inputs. 
                            ### 🤖 Then, the models will output the linguistic entity found in your prompt based on your selection!

                            Note: If you get an 'Error' in the gpt-3.5 model, check the following:
                            - Check that you entered your key correctly without any extra characters.
                            - If you used a free key, it means you exceeded your quota from the free API Key.
                            """)
                
                gr.Markdown("""
                            ### 📊 Linguistic Complexity

                            - We use existing tool, [LFTK](https://github.com/brucewlee/lftk?tab=readme-ov-file), to estimate the liguistic complexity of input sentences.
                            - For more information regarding the meanings of each feature keyword, please reference their documentation [here](https://docs.google.com/spreadsheets/d/1uXtQ1ah0OL9cmHp2Hey0QcHb4bifJcQFLvYlVIAWWwQ/edit#gid=693915416).
                            """)

            # Inputs
            ling_ents_prompt = gr.Textbox(show_label=False, placeholder="Write a prompt here")
            with gr.Row():
                # Will activate after getting API key
                have_key2 = gr.Dropdown(["Yes", "No"], label="Do you own an API Key?", scale=0.5)
                ling_ents_apikey_input = gr.Textbox(label="Open AI Key", placeholder="Enter your OpenAI key here", type="password")
            linguistic_entities = gr.Dropdown(["Noun", "Determiner", "Noun phrase", "Verb phrase", "Dependent clause", "T-units"], label="Linguistic Entity", allow_custom_value=True, info="If your choice is not included in the options, please type your own.")
            ling_ents_btn = gr.Button(value="Submit")

            # Outputs

            user_prompt_1 = gr.Textbox(label="Original prompt")
            linguistic_features_textbox = gr.Textbox(label="Linguistic Complexity", disabled=True)

            with gr.Row():
                gpt_ling_ents_chatbot = gr.Chatbot(label="gpt-3.5")
                llama_ling_ents_chatbot = gr.Chatbot(label="llama-7b")
                vicuna_ling_ents_chatbot = gr.Chatbot(label="vicuna-7b")
            # clear = gr.ClearButton(components=[ling_ents_prompt, ling_ents_apikey_input, have_key2, linguistic_entities,
            #                                    vicuna_ling_ents_chatbot, llama_ling_ents_chatbot, gpt_ling_ents_chatbot,])
                                               
            # Event Handler for API Key
            ling_ents_btn.click(update_api_key, inputs=ling_ents_apikey_input)

            def update_textbox(prompt):
                return prompt
                        
            ling_ents_btn.click(fn=update_textbox, inputs=ling_ents_prompt, outputs=user_prompt_1, api_name="ling_ents_btn")

            # Show features from LFTK 
            ling_ents_btn.click(linguistic_features_fn, inputs=[ling_ents_prompt], outputs=[linguistic_features_textbox])

            # Event Handler for GPT 3.5 Chatbot
            ling_ents_btn.click(gpt_respond, inputs=[have_key2, linguistic_entities, ling_ents_prompt, gpt_ling_ents_chatbot], 
                                outputs=[ling_ents_prompt, gpt_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 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])

        with gr.Tab("CoreNLP"):
            with gr.Row():
                gr.Markdown("""
                            ## 📜 Step-By-Step Instructions

                            - Enter a sentence for three models to process (Vicuna-7b, LLaMA-7b and GPT-3.5).
                            - If you own an OpenAI API key, select 'Yes' in the dropdown. If you don't own one, select 'No'.
                                - If you selected 'Yes', enter your OpenAI API Key [Link to your OpenAI keys](https://platform.openai.com/api-keys).
                                - If you selected 'No', leave the 'OpenAI Key' field blank and continue with the rest.
                            - Select a Task from the Dropdown.
                            - Select a Linguistic Entity from the Dropdown or enter a custom one.
                            - Click 'Submit' to send your inputs to the models.
                            - To submit a new prompt, repeat all the steps above and click 'Submit' again. Your new prompt should appear on the top of previous ones.

                            ### ⏳ After you click 'Submit', the models will take a couple seconds to process your inputs. 
                            ### 🤖 Then, the models will output the POS Tagging or Chunking in your prompt with three different strategies based on your selections!

                            Note: If you get an 'Error' in the gpt-3.5 model, check the following:
                            - Check that you entered your key correctly without any extra characters.
                            - If you used a free key, it means you exceeded your quota from the free API Key.
                            """)
                
                with gr.Column():
                    gr.Markdown("""
                            ### 📊 Linguistic Complexity

                            - We use existing tool, [LFTK](https://github.com/brucewlee/lftk?tab=readme-ov-file), to estimate the liguistic complexity of input sentences.
                            - For more information regarding the meanings of each feature keyword, please reference their documentation [here](https://docs.google.com/spreadsheets/d/1uXtQ1ah0OL9cmHp2Hey0QcHb4bifJcQFLvYlVIAWWwQ/edit#gid=693915416).
                            """)
                    
                    gr.Markdown("""
                                ### 🛠️ How each Strategy works

                                - Strategy 1 - QA-Based Prompting
                                    - The model is prompted with a question-answer format. The input consists of a question, and the model generates a response based on the understanding of the question and its knowledge.
                                - Strategy 2 - Instruction-Based Prompting
                                    - Involves providing the model with explicit instructions on how to generate a response. Instead of relying solely on context or previous knowledge, the instructions guide the model in generating content that aligns with specific criteria.
                                - Strategy 3 - Structured Prompting
                                    - Involves presenting information to the model in a structured format, often with defined sections or categories. The model then generates responses following the given structure. 
                                """)
            
            # Inputs
            task_prompt = gr.Textbox(show_label=False, placeholder="Write a prompt here")
            with gr.Row():
                have_key = gr.Dropdown(["Yes", "No"], label="Do you own an API Key?", scale=0.5)
                task_apikey_input = gr.Textbox(label="Open AI Key", placeholder="Enter your OpenAI key here", type="password", visible=True)
            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", allow_custom_value=True, info="If your choice is not included in the options, please type your own.")
            task_btn = gr.Button(value="Submit")

            # Outputs
            user_prompt_2 = gr.Textbox(label="Original prompt", )
            linguistic_features_textbox_2 = gr.Textbox(label="Linguistic Complexity", disabled=True, info="[Definitions for Complexity Indices](https://docs.google.com/spreadsheets/d/1uXtQ1ah0OL9cmHp2Hey0QcHb4bifJcQFLvYlVIAWWwQ/edit#gid=693915416)]")

            gr.Markdown("### Strategy 1 - QA-Based Prompting")
            strategy1 = gr.Markdown("S1", visible=False)
            with gr.Row():
                gpt_S1_chatbot = gr.Chatbot(label="gpt-3.5")
                llama_S1_chatbot = gr.Chatbot(label="llama-7b")
                vicuna_S1_chatbot = gr.Chatbot(label="vicuna-7b")
            gr.Markdown("### Strategy 2 - Instruction-Based Prompting")
            strategy2 = gr.Markdown("S2", visible=False)
            with gr.Row():
                gpt_S2_chatbot = gr.Chatbot(label="gpt-3.5")
                llama_S2_chatbot = gr.Chatbot(label="llama-7b")
                vicuna_S2_chatbot = gr.Chatbot(label="vicuna-7b")
            gr.Markdown("### Strategy 3 - Structured Prompting")
            strategy3 = gr.Markdown("S3", visible=False)
            with gr.Row():
                gpt_S3_chatbot = gr.Chatbot(label="gpt-3.5")
                llama_S3_chatbot = gr.Chatbot(label="llama-7b")
                vicuna_S3_chatbot = gr.Chatbot(label="vicuna-7b")
            # clear_all = gr.ClearButton(components=[task_prompt, task_apikey_input, have_key, task, task_linguistic_entities,
            #                                    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])
            
            # Event Handler for API Key
            task_btn.click(update_api_key, inputs=task_apikey_input)

            # Show user's original prompt  
            def update_textbox(prompt):
                return prompt
            
            task_btn.click(fn=update_textbox, inputs=task_prompt, outputs=user_prompt_2, api_name="task_btn")

            # Show features from LFTK 
            task_btn.click(linguistic_features_fn, inputs=[task_prompt], outputs=[linguistic_features_textbox_2])

            # 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=[have_key, strategy1, task, task_linguistic_entities, task_prompt, gpt_S1_chatbot], 
                                outputs=[task_prompt, gpt_S1_chatbot])
            task_btn.click(gpt_strategies_respond, inputs=[have_key, strategy2, task, task_linguistic_entities, task_prompt, gpt_S2_chatbot], 
                                outputs=[task_prompt, gpt_S2_chatbot])
            task_btn.click(gpt_strategies_respond, inputs=[have_key, strategy3, task, task_linguistic_entities, task_prompt, gpt_S3_chatbot], 
                                outputs=[task_prompt, gpt_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 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])

with gr.Blocks(theme=gr.themes.Soft()) as demo:
    gr.Markdown("""
                # Assessing the Articulate
                ## A Comparative Analysis of the Core Linguistic Knowledge in Large Language Models
                """)

    # load interface
    interface()
    
demo.launch()