app.py

import os
from huggingface_hub import login, create_repo, upload_folder
from autotrain import AutoTrainAdvanced
from datasets import load_dataset, Dataset
from dotenv import load_dotenv
import pandas as pd
import gradio as gr
from transformers import AutoModelForCausalLM, AutoTokenizer
import torch

load_dotenv()

hf_token = os.getenv('HF_TOKEN')
profile_name = os.getenv('HUGGINGFACE_PROFILE')

login(token=hf_token)

dataset1 = load_dataset("daqc/wikipedia-txt-spanish", split='train')
dataset2 = load_dataset("jorgeortizfuentes/universal_spanish_chilean_corpus", split='train')

df1 = pd.DataFrame(dataset1)
df2 = pd.DataFrame(dataset2)

combined_df = pd.concat([df1, df2], ignore_index=True)
combined_dataset = Dataset.from_pandas(combined_df)

task_type = "text-generation"
model_name = "meta-llama/Llama-3.2-1B"

config = {
    "task": task_type,
    "model": model_name,
    "train_data": combined_dataset,
    "output_dir": None,
    "epochs": 1,
    "learning_rate": 5e-5,
    "batch_size": 32,
    "fp16": True,
    "gradient_accumulation_steps": 4,
    "max_steps": 1,
}

model_repo_name = f"{profile_name}/llama-3-2-1b-text-generation"

create_repo(repo_id=model_repo_name, exist_ok=True, token=hf_token)

while True:
    trainer = AutoTrainAdvanced(config=config)
    trainer.train()

    upload_folder(
        folder_path="./output_model_llama",
        repo_id=model_repo_name,
        token=hf_token,
        repo_type="model"
    )

    print(f"Modelo subido correctamente a: https://huggingface.co./{model_repo_name}")
    print("Iteración de entrenamiento completada. Continuando con la siguiente...")

tokenizer = AutoTokenizer.from_pretrained(model_repo_name)
model = AutoModelForCausalLM.from_pretrained(model_repo_name)

def generate_text(input_text):
    inputs = tokenizer.encode(input_text, return_tensors='pt')
    with torch.no_grad():
        outputs = model.generate(inputs, max_length=50, num_return_sequences=1)
    return tokenizer.decode(outputs[0], skip_special_tokens=True)

iface = gr.Interface(fn=generate_text, inputs="text", outputs="text", title="Interacción con Llama 3.2", description="Escribe un texto y genera una respuesta.")
iface.launch()