app.py

import streamlit as st
from transformers import pipeline

# Initialize the pipeline for the fill-mask task
unmasker = pipeline('fill-mask', model='dsfsi/zabantu-bantu-250m')

# Sample sentences for different languages with placeholders for the masked word
sample_sentences = {
    'Zulu': "Le ndoda ithi izo____ ukudla.",  # Masked word for Zulu
    'Tshivenda': "Mufana uyo____ vhukuma.",  # Masked word for Tshivenda
    'Sepedi': "Mosadi o ____ pheka.",  # Masked word for Sepedi
    'Tswana': "Monna o ____ tsamaya.",  # Masked word for Tswana
    'Tsonga': "N'wana wa xisati u ____ ku tsaka."  # Masked word for Tsonga
}

# Function to perform fill-mask on sentences with the token '__' replaced
def fill_mask_for_languages(sentences):
    results = {}
    for language, sentence in sentences.items():
        # Replace the '____' placeholder with the model's mask token
        masked_sentence = sentence.replace('____', unmasker.tokenizer.mask_token)
        
        # Get predictions for the masked sentence
        unmasked = unmasker(masked_sentence)
        
        # Store the result for each language
        results[language] = unmasked
    return results

# Streamlit interface
st.title("Fill Mask for Multiple Languages | Zabantu-Bantu-250m")
st.write("This app predicts the missing word for sentences in Zulu, Tshivenda, Sepedi, Tswana, and Tsonga using a Zabantu BERT model.")

# Display the original sample sentences
st.write("### Sample sentences:")
for language, sentence in sample_sentences.items():
    st.write(f"**{language}**: {sentence}")

# User input for custom sentences
user_sentence = st.text_input("Enter your own sentence with a masked word (use '____'):", "Enter sentence here...")

# Add a submit button
if st.button("Submit"):
    # Prepare user input for prediction
    user_masked_sentence = user_sentence.replace('____', unmasker.tokenizer.mask_token)

    # Get predictions for the user input sentence
    user_predictions = unmasker(user_masked_sentence)

    # Display results for user input
    st.write("### Your Input:")
    st.write(f"Original sentence: {user_sentence}")
    st.write(f"Top prediction for the masked token: {user_predictions[0]['sequence']}")
    
    # Display results for sample sentences
    st.write("### Predictions for Sample Sentences:")
    for language, predictions in fill_mask_for_languages(sample_sentences).items():
        original_sentence = sample_sentences[language]
        predicted_sentence = predictions[0]['sequence']
        st.write(f"Original sentence ({language}): {original_sentence}")
        st.write(f"Top prediction for the masked token: {predicted_sentence}\n")
        st.write("=" * 80)

# Custom CSS styling for Streamlit elements
css = """
<style>
footer {display:none !important}

.stButton > button {
    background-color: #17152e;
    color: white;
    border: none;
    padding: 0.75em 2em;
    text-align: center;
    text-decoration: none;
    display: inline-block;
    font-size: 16px;
    margin: 4px 2px;
    cursor: pointer;
    border-radius: 12px;
    transition: background-color 0.3s ease;
}

.stButton > button:hover {
    background-color: #3c4a6b;
}

.stTextInput, .stTextArea {
    border: 1px solid #e6e6e6;
    padding: 0.75em;
    border-radius: 10px;
    font-size: 16px;
    width: 100%;
}

.stTextInput:focus, .stTextArea:focus {
    border-color: #17152e;
    outline: none;
    box-shadow: 0px 0px 5px rgba(23, 21, 46, 0.5);
}

div[data-testid="stMarkdownContainer"] p {
    font-size: 16px;
}

.stApp {
    padding: 2em;
    font-family: 'Poppins', sans-serif;
}
</style>
"""
st.markdown(css, unsafe_allow_html=True)