app.py

import random
from mtranslate import translate
import streamlit as st
from transformers import AutoTokenizer, AutoModelForMaskedLM, pipeline


LOGO = "https://raw.githubusercontent.com/nlp-en-es/assets/main/logo.png"

MODELS = {
    "RoBERTa Base Gaussian Seq Len 512": {
        "url": "bertin-project/bertin-base-gaussian-exp-512seqlen"
    },
    "RoBERTa Base Gaussian Seq Len 128": {
        "url": "bertin-project/bertin-base-gaussian"
    },
    "RoBERTa Base Random Seq Len 128": {
        "url": "bertin-project/bertin-base-random"
    },
}

PROMPT_LIST = [
    "Fui a la librería a comprar un <mask>.",
    "¡Qué buen <mask> hace hoy!",
    "Hoy empiezan las vacaciones así que vamos a la <mask>.",
    "Mi color favorito es el <mask>.",
    "Voy a <mask> porque estoy muy cansada.",
    "Mañana vienen mis amigos de <mask>.",
    "¿Te apetece venir a <mask> conmigo?",
    "En verano hace mucho <mask>.",
    "En el bosque había <mask>.",
    "El ministro dijo que <mask> los impuestos.",
    "Si no estuviera afónica, <mask> esa canción.",
]


@st.cache(show_spinner=False, persist=True)
def load_model(masked_text, model_url):
    model = AutoModelForMaskedLM.from_pretrained(model_url)
    tokenizer = AutoTokenizer.from_pretrained(model_url)
    nlp = pipeline("fill-mask", model=model, tokenizer=tokenizer)
    result = nlp(masked_text)
    return result


# Page
st.set_page_config(page_title="BERTIN Demo", page_icon=LOGO)
st.title("BERTIN")

#Sidebar
st.sidebar.image(LOGO)

# Body
st.markdown(
    """
    BERTIN is a series of BERT-based models for Spanish.
    
    The models are trained with Flax and using TPUs sponsored by Google since this is part of the
    [Flax/Jax Community Week](https://discuss.huggingface.co/t/open-to-the-community-community-week-using-jax-flax-for-nlp-cv/7104)
    organised by HuggingFace.
    
    All models are variations of **RoBERTa-base** trained from scratch in **Spanish** using the **mc4 dataset**. 
    We reduced the dataset size to 50 million documents to keep training times shorter, and also to be able to bias training examples based on their perplexity.
    
    The idea is to favour examples with perplexities that are neither too small (short, repetitive texts) or too long (potentially poor quality).
    * **Random** sampling simply takes documents at random to reduce the dataset size.
    * **Gaussian** rejects documents with a higher probability for lower and larger perplexities, based on a Gaussian function.
    
    The first models have been trained (250.000 steps) on sequence length 128, and training for Gaussian changed to sequence length 512 for the last 25.000 training steps.
    """
)

model_name = st.selectbox("Model", list(MODELS.keys()))
model_url = MODELS[model_name]["url"]

prompt = st.selectbox("Prompt", ["Random", "Custom"])
if prompt == "Custom":
    prompt_box = "Enter your masked text here..."
else:
    prompt_box = random.choice(PROMPT_LIST)
text = st.text_area("Enter text", prompt_box)

if st.button("Fill the mask"):
    with st.spinner(text="Filling the mask..."):
        st.subheader("Result")
        result = load_model(text, model_url)
        result_sequence = result[0]["sequence"]
        st.write(result_sequence)
        st.write("_English_ _translation:_", translate(result_sequence, "en", "es"))
        st.write(result)

st.markdown(
    """
    ### Team members
    - Eduardo González ([edugp](https://huggingface.co/edugp))
    - Javier de la Rosa ([versae](https://huggingface.co/versae))
    - Manu Romero ([mrm8488](https://huggingface.co/mrm8488))
    - María Grandury ([mariagrandury](https://huggingface.co/mariagrandury))
    - Pablo González de Prado ([Pablogps](https://huggingface.co/Pablogps))
    - Paulo Villegas ([paulo](https://huggingface.co/paulo))
        
    ### More information
    You can find more information about these models
    [here](https://huggingface.co/bertin-project/bertin-roberta-base-spanish).
    """
)