app.py

import streamlit as st
from transformers import pipeline

unmasker = pipeline('fill-mask', model='dsfsi/zabantu-xlm-roberta')

st.set_page_config(layout="wide")

def fill_mask(sentences):
    results = {}
    warnings = []
    for language, sentence in sentences.items():
        if "<mask>" in sentence:
            masked_sentence = sentence.replace('<mask>', unmasker.tokenizer.mask_token)
            unmasked = unmasker(masked_sentence)
            results[language] = unmasked
        else:
            warnings.append(f"Warning: No <mask> token found in sentence: {sentence}")
    return results, warnings

def replace_mask(sentence, predicted_word):
    return sentence.replace("<mask>", f"**{predicted_word}**")

st.title("Fill Mask | Zabantu-XLM-Roberta")
st.markdown("Zabantu-XLMR refers to a fleet of models trained on different combinations of South African Bantu languages.")

col1, col2 = st.columns(2)

if 'text_input' not in st.session_state:
    st.session_state['text_input'] = ""

if 'warnings' not in st.session_state:
    st.session_state['warnings'] = []

language_options = ['Zulu', 'Tshivenda', 'Sepedi', 'Tswana', 'Tsonga']

with col1:
    with st.container():
        st.markdown("### Input :clipboard:")
        
        input_sentences = {}
        for i in range(5):
            language = st.selectbox(f"Select language for sentence {i+1}:", language_options, key=f'language_{i}')
            sentence = st.text_area(f"Enter sentence for {language} (with <mask>):", key=f'text_input_{i}')
            if sentence:
                input_sentences[language.lower()] = sentence

        button1, button2, _ = st.columns([2, 2, 4])
        with button1:
            if st.button("Test Example"):
                sample_sentence = {
                    'zulu': "Le ndoda ithi izo <mask> ukudla.",
                    'tshivenda': "Mufana uyo <mask> vhukuma.",
                    'sepedi': "Mosadi o <mask> pheka.",
                    'tswana': "Monna o <mask> tsamaya.",
                    'tsonga': "N'wana wa xisati u <mask> ku tsaka."
                }
                input_sentences = sample_sentence
                result, warnings = fill_mask(input_sentences)
        
        with button2:
            if st.button("Submit"):
                result, warnings = fill_mask(input_sentences)
                st.session_state['warnings'] = warnings

        if st.session_state['warnings']:
            for warning in st.session_state['warnings']:
                st.warning(warning)

        st.markdown("### Example")
        st.code({
            'zulu': "Le ndoda ithi izo <mask> ukudla.",
            'tshivenda': "Mufana uyo <mask> vhukuma.",
            'sepedi': "Mosadi o <mask> pheka.",
            'tswana': "Monna o <mask> tsamaya.",
            'tsonga': "N'wana wa xisati u <mask> ku tsaka."
        }, wrap_lines=True)

with col2:
    with st.container():
        st.markdown("### Output :bar_chart:")
        if input_sentences:
            for language, sentence in input_sentences.items():
                masked_sentence = sentence.replace('<mask>', unmasker.tokenizer.mask_token)
                predictions = unmasker(masked_sentence)

                if predictions:
                    top_prediction = predictions[0]
                    predicted_word = top_prediction['token_str']
                    score = top_prediction['score'] * 100

                    st.markdown(f"""
                    <div class="bar">
                        <div class="bar-fill" style="width: {score}%;"></div>
                    </div>
                    <div class="container">
                        <div style="align-items: left;">{predicted_word} ({language})</div>
                        <div style="align-items: right;">{score:.2f}%</div>
                    </div>
                    """, unsafe_allow_html=True)

if 'predictions' in locals():  
    if result:
        for language, language_predictions in result.items(): 
            original_sentence = input_sentences[language]
            predicted_sentence = replace_mask(original_sentence, language_predictions[0]['token_str'])
            st.write(f"{language}: {predicted_sentence}\n")

css = """
<style>
footer {display:none !important;}

.container {
    display: flex;
    justify-content: space-between;
    align-items: center;
    margin-bottom: 5px;
    width: 100%;
}
.bar {
    background-color: #e6e6e6;
    border-radius: 12px;
    overflow: hidden;
    margin-right: 10px;
    height: 5px;
}
.bar-fill {
    background-color: #17152e;
    height: 100%;
    border-radius: 12px;
}
</style>
"""
st.markdown(css, unsafe_allow_html=True)