Spaces:
Sleeping
Sleeping
| 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") | |
| 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(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. Output any <{tab_name}> in the following sentence one per line.''', | |
| user_prompt = formatted_prompt) | |
| chat_history.append((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 one per line: "{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'''Output any {tab_name} in the following sentence one per line: "{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(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'''Generate the output only for the assistant. Output any {task_ling_ent} in the following sentence one per line: "{message}"''' | |
| elif (strategy == "S2"): | |
| formatted_system_prompt = f'''POS tag the following sentence using Universal POS tag set: "{message}"''' | |
| elif (strategy == "S3"): | |
| formatted_system_prompt = f'''POS tag the following sentence using Universal POS tag set: "{message}"''' | |
| elif (task_name == "Chunking"): | |
| if (strategy == "S1"): | |
| formatted_system_prompt = f'''Generate the output only for the assistant. Output any {task_ling_ent} in the following sentence one per line: "{message}"''' | |
| elif (strategy == "S2"): | |
| formatted_system_prompt = f'''Chunk the following sentence in CoNLL 2000 format with BIO tags: "{message}"''' | |
| elif (strategy == "S3"): | |
| formatted_system_prompt = f'''Chunk the following sentence in CoNLL 2000 format with BIO tags: "{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'''Output any {task_ling_ent} in the following sentence one per line: "{message}"''' | |
| elif (strategy == "S2"): | |
| formatted_prompt = f'''POS tag the following sentence using Universal POS tag set: "{message}"''' | |
| elif (strategy == "S3"): | |
| formatted_prompt = f'''POS tag the following sentence using Universal POS tag set: "{message}"''' | |
| elif (task_name == "Chunking"): | |
| if (strategy == "S1"): | |
| formatted_prompt = f'''Output any {task_ling_ent} in the following sentence one per line: "{message}"''' | |
| elif (strategy == "S2"): | |
| formatted_prompt = f'''Chunk the following sentence in CoNLL 2000 format with BIO tags: "{message}"''' | |
| elif (strategy == "S3"): | |
| formatted_prompt = f'''Chunk the following sentence in CoNLL 2000 format with BIO tags: "{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'''Output any {task_ling_ent} in the following sentence one per line: "{message}"''' | |
| elif (strategy == "S2"): | |
| formatted_prompt = f'''POS tag the following sentence using Universal POS tag set: "{message}"''' | |
| elif (strategy == "S3"): | |
| formatted_prompt = f'''POS tag the following sentence using Universal POS tag set: "{message}"''' | |
| elif (task_name == "Chunking"): | |
| if (strategy == "S1"): | |
| formatted_prompt = f'''Output any {task_ling_ent} in the following sentence one per line: "{message}"''' | |
| elif (strategy == "S2"): | |
| formatted_prompt = f'''Chunk the following sentence in CoNLL 2000 format with BIO tags: "{message}"''' | |
| elif (strategy == "S3"): | |
| formatted_prompt = f'''Chunk the following sentence in CoNLL 2000 format with BIO tags: "{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(): | |
| with gr.Tab("Linguistic Entities"): | |
| 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. | |
| - Click 'Submit' to send your inputs to the models. | |
| - To enter a new prompt, scroll to the bottom and click 'Clear' to start again. | |
| ### ⏳ 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. | |
| """) | |
| # Inputs | |
| task_prompt = gr.Textbox(show_label=False, placeholder="Write a prompt and press enter") | |
| 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") | |
| task_btn = gr.Button(value="Submit") | |
| # Outputs | |
| 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) | |
| # 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]) | |
| # 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]) | |
| 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() | |