app.py

import json
import time
from typing import List, Tuple

import gradio as gr
import spaces
import torch
from transformers import AutoModelForCausalLM, AutoTokenizer

MODEL_ID = "Qwen/Qwen3-0.6B"
MAX_TEXT_CHARS = 4000

tokenizer = AutoTokenizer.from_pretrained(MODEL_ID)
model = AutoModelForCausalLM.from_pretrained(
    MODEL_ID,
    torch_dtype=torch.float16,
    device_map="auto",
)

PRESET_LABELS = {
    "Sentiment": "positive, negative, neutral",
    "Topic": "tech, business, sports, science, politics",
    "Intent": "question, statement, request, complaint",
    "Tone": "formal, casual, urgent, friendly",
}


def normalize_labels(raw_labels: str) -> List[str]:
    labels = [label.strip() for label in raw_labels.split(",") if label.strip()]
    unique_labels = []
    seen = set()
    for label in labels:
        key = label.lower()
        if key not in seen:
            seen.add(key)
            unique_labels.append(label)
    return unique_labels


def truncate_text(text: str) -> Tuple[str, bool]:
    text = (text or "").strip()
    if len(text) <= MAX_TEXT_CHARS:
        return text, False
    return text[:MAX_TEXT_CHARS], True


def build_bar_chart(labels: List[str], scores: List[float]) -> str:
    if not labels:
        return "<p>No labels to display.</p>"

    rows = []
    max_score = max(scores) if scores else 1.0
    for label, score in zip(labels, scores):
        pct = score * 100
        width = (score / max_score) * 100 if max_score > 0 else 0
        rows.append(
            f"""
            <div style='margin: 8px 0;'>
              <div style='display:flex; justify-content:space-between; font-size:14px;'>
                <span><b>{label}</b></span>
                <span>{pct:.2f}%</span>
              </div>
              <div style='background:#e5e7eb; border-radius:8px; overflow:hidden;'>
                <div style='width:{width:.2f}%; background:#2563eb; color:white; padding:4px 8px; font-size:12px;'>
                  {pct:.2f}%
                </div>
              </div>
            </div>
            """
        )

    return "<div>" + "\n".join(rows) + "</div>"


def apply_preset(preset_name: str) -> str:
    return PRESET_LABELS.get(preset_name, "")


def build_classification_prompt(text: str, labels: List[str], multi_label: bool) -> str:
    labels_str = ", ".join(f'"{l}"' for l in labels)
    mode_instruction = (
        "Multiple labels can apply simultaneously. For each label, assign a confidence score between 0 and 1."
        if multi_label
        else "Choose the single best label. Assign confidence scores that sum to 1."
    )

    return (
        f"Classify the following text into these categories: {labels_str}\n\n"
        f"{mode_instruction}\n\n"
        f"Text: \"{text}\"\n\n"
        f"Respond with ONLY a JSON object mapping each label to its confidence score. "
        f"Example: {{{', '.join(f'\"{l}\": 0.5' for l in labels[:2])}}}\n"
        f"JSON:"
    )


def parse_scores(output: str, labels: List[str]) -> dict:
    """Extract label scores from model output, with fallback parsing."""
    # Try to find JSON in the output
    output = output.strip()

    # Find the first { and last }
    start = output.find("{")
    end = output.rfind("}")
    if start != -1 and end != -1 and end > start:
        json_str = output[start : end + 1]
        try:
            parsed = json.loads(json_str)
            scores = {}
            for label in labels:
                # Try exact match, then case-insensitive
                if label in parsed:
                    scores[label] = float(parsed[label])
                else:
                    lower_map = {k.lower(): v for k, v in parsed.items()}
                    scores[label] = float(lower_map.get(label.lower(), 0.0))
            return scores
        except (json.JSONDecodeError, ValueError):
            pass

    # Fallback: equal scores
    return {label: 1.0 / len(labels) for label in labels}


@spaces.GPU
def run_classification(text: str, candidate_labels: str, multi_label: bool):
    clean_text, was_truncated = truncate_text(text)
    labels = normalize_labels(candidate_labels)

    if not clean_text:
        raise gr.Error("Please enter text to classify.")
    if len(labels) < 2:
        raise gr.Error("Please provide at least 2 labels, separated by commas.")

    prompt = build_classification_prompt(clean_text, labels, multi_label)

    messages = [
        {"role": "system", "content": "You are a precise text classifier. Respond only with valid JSON."},
        {"role": "user", "content": prompt},
    ]

    input_text = tokenizer.apply_chat_template(
        messages, tokenize=False, add_generation_prompt=True,
        enable_thinking=False,
    )

    start = time.perf_counter()

    inputs = tokenizer(input_text, return_tensors="pt").to(model.device)
    with torch.no_grad():
        outputs = model.generate(
            **inputs,
            max_new_tokens=256,
            temperature=0.1,
            do_sample=True,
            top_p=0.9,
        )

    generated = tokenizer.decode(outputs[0][inputs["input_ids"].shape[1]:], skip_special_tokens=True)
    elapsed = time.perf_counter() - start

    scores = parse_scores(generated, labels)

    # Normalize scores
    total = sum(scores.values())
    if total > 0:
        scores = {k: v / total for k, v in scores.items()}
    else:
        scores = {k: 1.0 / len(labels) for k in labels}

    sorted_pairs = sorted(scores.items(), key=lambda x: x[1], reverse=True)
    sorted_labels = [x[0] for x in sorted_pairs]
    sorted_scores = [x[1] for x in sorted_pairs]

    chart_html = build_bar_chart(sorted_labels, sorted_scores)

    top_label = sorted_labels[0]
    top_score = sorted_scores[0] * 100
    truncation_note = (
        f" Input was truncated to {MAX_TEXT_CHARS} characters for stable inference."
        if was_truncated
        else ""
    )

    summary = (
        f"Top prediction: {top_label} ({top_score:.2f}%). "
        f"Model: Qwen3-0.6B. "
        f"Mode: {'multi-label' if multi_label else 'single-label'}. "
        f"Inference time: {elapsed:.3f}s.{truncation_note}"
    )

    return chart_html, summary


with gr.Blocks(theme=gr.themes.Soft(), title="Zero-Shot Text Classifier") as demo:
    gr.Markdown("# Zero-Shot Text Classifier")
    gr.Markdown(
        "Classify any text into custom labels using **Qwen3-0.6B** with zero-shot instruction prompting. "
        "No fine-tuning needed: define your own categories and classify instantly."
    )

    with gr.Row():
        preset = gr.Dropdown(
            choices=list(PRESET_LABELS.keys()),
            value="Sentiment",
            label="Preset Label Sets",
            info="Pick a preset to auto-fill candidate labels.",
        )
        apply_btn = gr.Button("Apply Preset")

    text_input = gr.Textbox(
        label="Text to Classify",
        placeholder="Paste a sentence, paragraph, or document...",
        lines=8,
    )

    labels_input = gr.Textbox(
        label="Candidate Labels (comma-separated)",
        value=PRESET_LABELS["Sentiment"],
        placeholder="positive, negative, neutral",
        lines=2,
    )

    multi_label_input = gr.Checkbox(
        label="Multi-label mode",
        value=False,
        info="If enabled, multiple labels can be true at the same time.",
    )

    classify_btn = gr.Button("Classify", variant="primary")

    chart_output = gr.HTML(label="Label Scores")
    summary_output = gr.Textbox(label="Summary", interactive=False)

    apply_btn.click(apply_preset, inputs=preset, outputs=labels_input)
    classify_btn.click(
        run_classification,
        inputs=[text_input, labels_input, multi_label_input],
        outputs=[chart_output, summary_output],
    )

    gr.Markdown(
        "Built by [Xavier Fuentes](https://huggingface.co/xavier-fuentes) @ "
        "[AI Enablement Academy](https://enablement.academy) | "
        "[Buy me a coffee ☕](https://ko-fi.com/xavierfuentes)"
    )


if __name__ == "__main__":
    demo.queue().launch()