app.py

# gradio app for the LLM model --> use the retr environment
# Run the script and open the link in the browser.

import os
import pandas as pd
import datasets
import gradio as gr
import torch
from transformers import AutoTokenizer, AutoModelForCausalLM, pipeline

# training from scratch with latbert tokenizer
CHECKPOINT_PATH= 'scratch_2-nodes_tokenizer_latbert-original_packing_fcocchi/'
CHECKPOINT_PATH= 'itserr/scratch_2-nodes_tokenizer_latbert-original_packing_fcocchi'

print(f"Loading model from: {CHECKPOINT_PATH}")
tokenizer = AutoTokenizer.from_pretrained(CHECKPOINT_PATH, token=os.environ['HF_TOKEN_READ'])
model = AutoModelForCausalLM.from_pretrained(CHECKPOINT_PATH, token=os.environ['HF_TOKEN_READ'])

preference_dataset_name= "itserr/latin_gpt_preferences"
global dataset_hf
dataset_hf = datasets.load_dataset(preference_dataset_name, token=os.environ['HF_TOKEN_READ'], download_mode='force_redownload')
dataset_hf = dataset_hf['train'].to_pandas()
print(dataset_hf.shape)

description="""
This is a Latin Language Model (LLM) based on GPT-2 and it was trained on a large corpus of Latin texts and can generate text in Latin. \n
Demo instructions:
- Enter a prompt in Latin in the Input Text box.
- Select the temperature value to control the randomness of the generated text (higher value produce a more creative and unstable answer).
- Click the 'Generate Text' button to trigger model generation.
- (Optional) insert a Feedback text in the box.
- Click the 'Like' or 'Dislike' button to judge the generation correctness. 
"""
# (L<sup>2</sup>) - Latin Language Model
title= "LatinGPT"
article= "hello world ..."
examples= ['Accidere ex una scintilla', 'Audacter calumniare,', 'Consolatium misero comites', 'Errare humanum est,', 'Excusatio non petita,']
logo_image= 'ITSERR_row_logo.png'

def generate_text(prompt, slider):
    if torch.cuda.is_available(): device = torch.device("cuda")      
    else: 
        device = torch.device("cpu")
        print("No GPU available")
    
    print("***** Generate *****")
    text_generator = pipeline("text-generation", model=model, tokenizer=tokenizer, device=device)
    #generated_text = text_generator(prompt, max_length=100)
    if slider is not None:
        slider = float(slider)
    generated_text = text_generator(prompt, max_length=50, do_sample=True, temperature=slider, repetition_penalty=2.0, truncation=True, min_new_tokens=5)
    return generated_text[0]['generated_text']

# Function to handle user preferences
def handle_preference(preference, input, output, feedback, temp_value):
    """
    Format values stored in preferences:
        - input text
        - output generated text
        - user feedback
        - float temperature value
    """
    # first time staring from a csv file (edited the present one), then work with parquet file
    # input_text,generated_text,feedback,temperature,like,dislike,count_like,count_dislike
    global dataset_hf
    if input == output:
        output_tuple= ("", "")
    else:
        output_tuple= (input, output.split(input)[-1])
    if preference == "like":
        dislike=0
        like=1
        count_like= dataset_hf.iloc[-1]['count_like']
        count_dislike= dataset_hf.iloc[-1]['count_dislike']
        if output_tuple[1] != "" :
            count_like= dataset_hf.iloc[-1]['count_like'] + 1

    elif preference == "dislike":
        dislike=1
        like=0
        count_like= dataset_hf.iloc[-1]['count_like']
        count_dislike= dataset_hf.iloc[-1]['count_dislike']
        if output_tuple[1] != "" :
            count_dislike= dataset_hf.iloc[-1]['count_dislike'] + 1

    inp_text= output_tuple[0]
    out_text= output_tuple[1]
    new_data = pd.DataFrame({'input_text': inp_text, 'generated_text': out_text, 'feedback': feedback,
                              'temperature': float(temp_value), 'like': like, 'dislike': dislike, 
                              'count_like': count_like, 'count_dislike': count_dislike}, index=[0])
    dataset_hf = pd.concat([dataset_hf, new_data], ignore_index=True)
    hf_dataset = datasets.Dataset.from_pandas(dataset_hf)
    dataset_dict = datasets.DatasetDict({"train": hf_dataset})
    dataset_dict.push_to_hub(preference_dataset_name, token=os.environ['HF_TOKEN_WRITE'])
    
    # print dataset statistics
    print(f"Admin log: like: {count_like} and dislike: {count_dislike}")
    return f"You select '{preference}' as answer of the model generation. Thank you for your time!"

custom_css = """
#logo {
    display: block;
    margin-left: auto;
    margin-right: auto;
    width: 280px;
    height: 140px;
}
"""

with gr.Blocks(css=custom_css) as demo:
    gr.Image(logo_image, elem_id="logo")
    gr.Markdown(f"<h1 style='text-align: center;'>{title}</h1>")
    gr.Markdown(description)
    
    with gr.Row():
        with gr.Column():
            input_text = gr.Textbox(lines=5, placeholder="Enter latin text here...", label="Input Text")
        with gr.Column():
            output_text = gr.Textbox(lines=5, placeholder="Output text will appear here...", label="Output Text")

    gr.Examples(examples=examples, inputs=input_text, cache_examples=True, fn=generate_text, outputs=output_text) # , cache_examples="true"
    temperature_slider = gr.Slider(minimum=0.1, maximum=5.0, step=0.1, value=1.0, label="Temperature")
    
    clean_button = gr.Button("Generate Text")
    clean_button.click(fn=generate_text, inputs=[input_text, temperature_slider], outputs=output_text)
    feedback_output = gr.Textbox(lines=1, placeholder="If you want to provide a feedback, please fill this box ...", label="Feedback")

    with gr.Row():
        like_button = gr.Button("Like")
        dislike_button = gr.Button("Dislike")

    button_output = gr.Textbox(lines=1, placeholder="Please submit your choice", label="Latin Language Model Demo")
    like_button.click(fn=lambda x,y,z,v: handle_preference("like", x, y, z, v), inputs=[input_text, output_text, feedback_output, temperature_slider], outputs=button_output)
    dislike_button.click(fn=lambda x,y,z,v: handle_preference("dislike", x, y, z, v), inputs=[input_text, output_text, feedback_output, temperature_slider], outputs=button_output)
    #gr.Markdown(article)

demo.launch(share=True)