app.py

import gradio as gr
from transformers import pipeline, AutoTokenizer

# --- Configuration ---
# IMPORTANT: Replace with the actual HF Hub repo name where you pushed your model
MODEL_NAME = "Bur3hani/karani-afro-xlmr-base-finetuned-swahili"
# --- End Configuration ---

# Load the tokenizer to access the mask token easily
try:
    tokenizer = AutoTokenizer.from_pretrained(MODEL_NAME)
    mask_token = tokenizer.mask_token if tokenizer.mask_token else "[MASK]" # Default if somehow missing
except Exception as e:
    print(f"Warning: Could not load tokenizer separately. Using default mask '[MASK]'. Error: {e}")
    mask_token = "[MASK]"


# Load the fill-mask pipeline using your fine-tuned model
try:
    mask_filler = pipeline("fill-mask", model=MODEL_NAME, tokenizer=MODEL_NAME)
    print(f"Pipeline loaded successfully for model: {MODEL_NAME}")
except Exception as e:
    print(f"Error loading pipeline: {e}")
    # Handle error, maybe disable the interface or show an error message
    mask_filler = None


# Define the function that Gradio will call
def fill_mask_kiswahili(sentence_with_mask):
    if mask_filler is None:
        return "ERROR: Model pipeline could not be loaded."
    if mask_token not in sentence_with_mask:
        return f"ERROR: Input sentence must contain the mask token: {mask_token}"

    try:
        predictions = mask_filler(sentence_with_mask, top_k=5)

        # Format the output
        output_text = f"Input: {sentence_with_mask}\n\nPredictions:\n"
        output_text += "-------------------\n"
        for i, pred in enumerate(predictions):
            token_str = pred['token_str'].replace(' ', '').strip() # Clean token
            sequence_reconstructed = sentence_with_mask.replace(mask_token, token_str)
            score = pred['score']
            output_text += f"{i+1}. {token_str} (Score: {score:.4f})\n   -> '{sequence_reconstructed}'\n"
        return output_text

    except Exception as e:
        return f"An error occurred during prediction: {e}"


# Create the Gradio interface
iface = gr.Interface(
    fn=fill_mask_kiswahili,
    inputs=gr.Textbox(lines=2, label="Sentence in Kiswahili (use '[MASK]' for the blank)", placeholder=f"Example: Leo hali ya hewa ni {mask_token} sana."),
    outputs=gr.Textbox(label="Top 5 Predictions"),
    title="Karani Kiswahili Assistant (Fill-Mask Demo)",
    description="Demo of a Kiswahili language model fine-tuned using Afro-XLMR-Base. Type a sentence in Kiswahili and use '[MASK]' where you want the model to predict a word.",
    examples=[
        [f"Mradi huu wa {mask_token} utasaidia sana."],
        [f"Alikwenda sokoni kununua {mask_token}."],
        [f"Ninasoma {mask_token} kwa makini."],
    ],
    allow_flagging="never" # Disable flagging for simplicity
)

# Launch the interface (when running app.py in the Space)
if __name__ == "__main__":
    iface.launch()