app.py

import gradio as gr
import torch
from transformers import AutoTokenizer, AutoModelForTokenClassification


# ============================================================
# BalPOS - Balochi Part-of-Speech Tagger
# CPU Hugging Face Space
# ============================================================

MODEL_NAME = "shahbakhsh/BalPOS"
DEVICE = torch.device("cpu")

print("=" * 60)
print("BalPOS - Balochi Part-of-Speech Tagger")
print("=" * 60)
print(f"Model: {MODEL_NAME}")
print("Device: CPU")
print("Loading model...")


# ============================================================
# LOAD MODEL ONCE
# ============================================================

tokenizer = AutoTokenizer.from_pretrained(MODEL_NAME)

model = AutoModelForTokenClassification.from_pretrained(MODEL_NAME)

model.to(DEVICE)
model.eval()

print("Model loaded successfully.")
print(f"Labels: {model.config.num_labels}")
print("=" * 60)


# ============================================================
# POS COLORS
# ============================================================

LABEL_COLORS = {
    "NOUN": "#93c5fd",
    "PROPN": "#a5b4fc",
    "PRON": "#c4b5fd",
    "VERB": "#86efac",
    "AUX": "#6ee7b7",
    "ADJ": "#fde68a",
    "ADV": "#fca5a5",
    "ADP": "#fdba74",
    "CCONJ": "#f9a8d4",
    "SCONJ": "#f0abfc",
    "DET": "#67e8f9",
    "NUM": "#d9f99d",
    "PART": "#e5e7eb",
    "PUNCT": "#d1d5db",
    "INTJ": "#fecaca",
    "X": "#e5e7eb",
}


# ============================================================
# POS TAGGING FUNCTION
# ============================================================

def tag_sentence(sentence):

    if not sentence or not sentence.strip():
        return (
            [],
            [],
            "⚠️ Please enter a Balochi sentence."
        )

    sentence = sentence.strip()
    words = sentence.split()

    # Keep CPU inference practical
    if len(words) > 128:
        return (
            [],
            [],
            "⚠️ Please enter a shorter sentence. "
            "Maximum supported length is 128 words."
        )

    try:

        # Preserve word boundaries
        inputs = tokenizer(
            words,
            is_split_into_words=True,
            return_tensors="pt",
            truncation=True,
            max_length=256
        )

        word_ids = inputs.word_ids(batch_index=0)

        inputs = {
            key: value.to(DEVICE)
            for key, value in inputs.items()
        }

        # CPU inference
        with torch.inference_mode():

            outputs = model(**inputs)

            logits = outputs.logits

            probabilities = torch.softmax(logits, dim=-1)

            predictions = logits.argmax(dim=-1)[0]

            confidences = probabilities.max(dim=-1).values[0]

        results = []
        table_rows = []
        seen_words = set()

        for token_index, word_id in enumerate(word_ids):

            if word_id is None:
                continue

            if word_id in seen_words:
                continue

            if word_id >= len(words):
                continue

            seen_words.add(word_id)

            predicted_id = predictions[token_index].item()

            label = model.config.id2label.get(
                predicted_id,
                str(predicted_id)
            )

            confidence = confidences[token_index].item() * 100

            word = words[word_id]

            results.append((word, label))

            table_rows.append([word, label, f"{confidence:.1f}%"])

        if not results:
            return (
                [],
                [],
                "⚠️ No predictions were produced."
            )

        average_confidence = sum(
            float(row[2].replace("%", ""))
            for row in table_rows
        ) / len(table_rows)

        summary = (
            f"**{len(results)} tokens** · "
            f"**{len(set(label for _, label in results))} POS categories** · "
            f"**{average_confidence:.1f}% average confidence**"
        )

        return (results, table_rows, summary)

    except Exception as error:

        print("Prediction error:", repr(error))

        return (
            [],
            [],
            "❌ An error occurred while processing the sentence."
        )


# ============================================================
# EXAMPLES
# ============================================================

EXAMPLES = [
    ["چہ وتی فلسفہ ءِ استاد ءَ آ اشکیتگ ات"],
    ["اے دوئیں دبستانانی گوناپ دیگءَ پرکے ھست۔"],
    ["اولی جست ءِ پسو نہ اِنت۔"],
    ["تہ اے پیمیں جست بے نہ بنت۔"],
]


# ============================================================
# CUSTOM CSS
# ============================================================

CUSTOM_CSS = """
.gradio-container {
    max-width: 1200px !important;
    margin: auto !important;
}

#hero {
    text-align: center;
    padding: 24px 12px 12px 12px;
}

#hero h1 {
    font-size: clamp(2rem, 5vw, 3.2rem);
    margin-bottom: 8px;
}

#balochi-input textarea {
    direction: rtl !important;
    text-align: right !important;
    font-size: 1.15rem !important;
    line-height: 2 !important;
}

#tag-button {
    min-height: 48px !important;
    font-weight: 600 !important;
}

.card {
    border-radius: 16px !important;
    padding: 16px !important;
}

#summary {
    text-align: center;
    padding: 10px;
}

@media (max-width: 768px) {
    .gradio-container {
        padding-left: 10px !important;
        padding-right: 10px !important;
    }

    #hero {
        padding-top: 12px;
    }

    #hero h1 {
        font-size: 2rem;
    }

    #balochi-input textarea {
        font-size: 1rem !important;
    }
}

@media (max-width: 430px) {
    #hero h1 {
        font-size: 1.7rem;
    }

    #hero p {
        font-size: 0.9rem;
    }
}

footer {
    display: none !important;
}
"""


# ============================================================
# BUILD GRADIO APP
# ============================================================

with gr.Blocks(title="BalPOS - Balochi POS Tagger") as demo:

    # --------------------------------------------------------
    # HEADER
    # --------------------------------------------------------
    with gr.Column(elem_id="hero"):

        gr.Markdown(
            """
# 🏔️ BalPOS

## Balochi Part-of-Speech Tagger
"""
        )

        gr.Markdown(
            """
Automatically predict Universal Part-of-Speech (UPOS)
tags for Balochi text.
"""
        )

    # --------------------------------------------------------
    # MAIN
    # --------------------------------------------------------
    with gr.Row():

        # INPUT
        with gr.Column(scale=1, elem_classes=["card"]):

            gr.Markdown("### ✍️ Enter Balochi Text")

            sentence_input = gr.Textbox(
                label="Balochi Sentence",
                placeholder="بلوچی جملہ ایتلاپ کن ...",
                lines=5,
                rtl=True,
                text_align="right",
                elem_id="balochi-input"
            )

            with gr.Row():

                clear_btn = gr.Button("Clear")

                submit_btn = gr.Button(
                    "🏷️ Tag Sentence",
                    variant="primary",
                    elem_id="tag-button"
                )

            gr.Examples(
                examples=EXAMPLES,
                inputs=sentence_input,
                label="Try an example"
            )

        # OUTPUT
        with gr.Column(scale=1, elem_classes=["card"]):

            gr.Markdown("### 🔎 POS Predictions")

            highlighted_output = gr.HighlightedText(
                label="Tagged Text",
                color_map=LABEL_COLORS,
                show_legend=True
            )

            summary_output = gr.Markdown(
                "Enter a sentence to begin.",
                elem_id="summary"
            )

    # --------------------------------------------------------
    # TABLE
    # --------------------------------------------------------
    with gr.Column(elem_classes=["card"]):

        gr.Markdown("### 📊 Token-Level Results")

        table_output = gr.Dataframe(
            headers=["Token", "UPOS", "Confidence"],
            datatype=["str", "str", "str"],
            interactive=False,
            wrap=True
        )

    # --------------------------------------------------------
    # MODEL INFORMATION
    # --------------------------------------------------------
    with gr.Row():

        with gr.Column(elem_classes=["card"]):

            gr.Markdown(
                """
### 🧠 Model Information

**Model:** `shahbakhsh/BalPOS`

**Task:** Balochi POS Tagging

**Backbone:** BalBERT

**UPOS labels:** 16

### Reported final evaluation

| Metric | Score |
|---|---:|
| Accuracy | **87.26%** |
| Balanced Accuracy | **78.81%** |
| Macro F1 | **78.99%** |
| Weighted F1 | **87.22%** |
| MCC | **0.8526** |
"""
            )

        with gr.Column(elem_classes=["card"]):

            gr.Markdown(
                """
### 🏷️ Supported UPOS Tags

`NOUN` · `PROPN` · `PRON` · `VERB`

`AUX` · `ADJ` · `ADV` · `ADP`

`CCONJ` · `SCONJ` · `DET` · `NUM`

`PART` · `PUNCT` · `INTJ` · `X`
"""
            )

  

    # ==========================================================
    # IMPORTANT: ALL EVENT HANDLERS MUST STAY INSIDE gr.Blocks
    # (this indentation is what the previous version was missing)
    # ==========================================================

    submit_btn.click(
        fn=tag_sentence,
        inputs=sentence_input,
        outputs=[highlighted_output, table_output, summary_output]
    )

    sentence_input.submit(
        fn=tag_sentence,
        inputs=sentence_input,
        outputs=[highlighted_output, table_output, summary_output]
    )

    clear_btn.click(
        fn=lambda: ("", [], [], "Enter a sentence to begin."),
        inputs=None,
        outputs=[sentence_input, highlighted_output, table_output, summary_output]
    )


# ============================================================
# LAUNCH
# ============================================================

if __name__ == "__main__":
    demo.launch(
        css=CUSTOM_CSS,
        theme=gr.themes.Soft(
            primary_hue="teal",
            neutral_hue="slate"
        )
    )