app.py

"""GigaCheck demo: classify text and localize AI-written spans.

Two tabs back two Mistral-7B models from the LLMTrace / GigaCheck project:
a binary human/AI classifier and a fine-grained AI-span detector.
"""

from __future__ import annotations

import gradio as gr
from loguru import logger

from models import classify_text, detect_intervals
from rendering import build_classifier_card, build_detector_card
from styles import CSS, THEME
from validation import CONFIG, validate_text

LANG_NOTE = (
    '<div class="gc-lang-note">⚠️ These models work with <b>English</b> and '
    f"<b>Russian</b> text only. Enter between {CONFIG.min_words} and "
    f"{CONFIG.max_words} words.</div>"
)

CLASSIFIER_EXAMPLES = [
    # human
    [
        "Netflix has laid off around 300 people across a number of departments. The company also laid off 150 workers back in May."
    ],
    # ai
    [
        "@GOP @PamBondi Maybe we should talk about Gordon Sondland, who had ZERO experience or qualifications to be an Ambassador to the EU but since he donated a million dollars to trump he got the job anyway. I'm outraged!"
    ],
    # ai
    [
        "В прекрасное утро, поразмыслив о том, как множество радостей приносит нам природа, я вспомнила о своей милой питомице, мадам Леске, и о ее необъяснимой страсти к дыне. \"Как же,\" подумала я, \"этот нежный плод вмещает в себе не только великолепие вкуса, но и целую кладезь полезных веществ\". И вдруг осенило меня: разве не заслуживает моя преданная спутница опробовать хоть крохотный кусочек этого плода?"
    ],
]

CLASSIFIER_EXAMPLE_LABELS = ["EN · human", "EN · AI", "RU · AI"]

DETECTOR_EXAMPLES = [
    [
       "The critic's review of the recent publication was scathing. The book failed miserably in portraying the harmful subjective discourses associated with the hegemony of the political system."

    ],
    [
        "Университет Шеффилд Холлем имеет в Шеффилде два кампуса. Один из них — это Городской кампус, который находится в центре города. Второй кампус, Коллегиальный, расположен на юго-западе Шеффилда. Оба кампуса предлагают современные учебные здания и ресурсы для студентов. История университета Шеффилд Холлем началась в 1843 году с основания Шеффилдской школы дизайна. В 1960-х годах несколько независимых колледжей (включая Школу дизайна) объединились в Шеффилдский Политехникум, с 1976 года — Шеффилдский городской политехникум, с 1992 года — Университет Шеффилд Холлем."
    ],
]

DETECTOR_EXAMPLE_LABELS = ["EN · example", "RU · example"]


def update_counter(text: str) -> str:
    """Render a live word counter, flagging out-of-bounds input in red.

    Args:
        text: Current textbox content.

    Returns:
        An HTML snippet showing the word count and any validation message.
    """
    result = validate_text(text)
    bad = (not result.ok) and result.word_count > 0
    css_class = "gc-counter gc-bad" if bad else "gc-counter"
    note = f" — {result.message}" if bad else ""
    return (
        f'<div class="{css_class}">{result.word_count} / '
        f"{CONFIG.max_words} words{note}</div>"
    )


def run_classifier(text: str) -> str:
    """Validate then classify the text, returning a result card.

    Args:
        text: User-supplied text.

    Returns:
        HTML for the classifier result card.

    Raises:
        gr.Error: If the text falls outside the configured word bounds.
    """
    result = validate_text(text)
    if not result.ok:
        raise gr.Error(result.message)
    label, p_human, p_ai = classify_text(text)
    logger.info("classifier: label={} p_human={:.3f}", label, p_human)
    return build_classifier_card(label, p_human, p_ai)


def run_detector(text: str, conf_threshold: float) -> str:
    """Validate then run AI-span detection, returning a result card.

    Args:
        text: User-supplied text.
        conf_threshold: Confidence threshold for keeping a span.

    Returns:
        HTML for the detector result card.

    Raises:
        gr.Error: If the text falls outside the configured word bounds.
    """
    result = validate_text(text)
    if not result.ok:
        raise gr.Error(result.message)
    intervals = detect_intervals(text, conf_threshold)
    logger.info("detector: {} interval(s) at thresh={}", len(intervals), conf_threshold)
    return build_detector_card(text, intervals)


def build_classifier_tab() -> None:
    """Build the classifier tab UI and wire its events."""
    gr.HTML(LANG_NOTE)
    text_in = gr.Textbox(
        label="Text",
        placeholder="Paste English or Russian text to classify…",
        lines=8,
    )
    counter = gr.HTML(update_counter(""))
    analyze_btn = gr.Button("Analyze", variant="primary")
    output = gr.HTML()
    gr.Examples(
        examples=CLASSIFIER_EXAMPLES,
        inputs=[text_in],
        label="Examples",
        example_labels=CLASSIFIER_EXAMPLE_LABELS,
    )

    text_in.change(update_counter, inputs=[text_in], outputs=[counter])
    analyze_btn.click(run_classifier, inputs=[text_in], outputs=[output])


def build_detector_tab() -> None:
    """Build the detector tab UI and wire its events."""
    gr.HTML(LANG_NOTE)
    text_in = gr.Textbox(
        label="Text",
        placeholder="Paste English or Russian text to scan for AI-written spans…",
        lines=8,
    )
    counter = gr.HTML(update_counter(""))
    conf = gr.Slider(
        label="Confidence threshold",
        minimum=0.0,
        maximum=1.0,
        value=CONFIG.default_conf_threshold,
        step=0.05,
    )
    detect_btn = gr.Button("Detect", variant="primary")
    output = gr.HTML()
    gr.Examples(
        examples=DETECTOR_EXAMPLES,
        inputs=[text_in],
        label="Examples",
        example_labels=DETECTOR_EXAMPLE_LABELS,
    )

    text_in.change(update_counter, inputs=[text_in], outputs=[counter])
    detect_btn.click(run_detector, inputs=[text_in, conf], outputs=[output])


def build_demo() -> gr.Blocks:
    """Assemble the full Gradio demo.

    Returns:
        The configured :class:`gradio.Blocks` application.
    """
    with gr.Blocks(theme=THEME, css=CSS, title="GigaCheck") as demo:
        gr.HTML(
            '<div id="gc-header"><h1>GigaCheck</h1>'
            "<p>Detect AI-generated text — binary classification and "
            "fine-grained span localization</p></div>"
        )
        with gr.Tabs():
            with gr.Tab("Classifier"):
                gr.Markdown(
                    "Binary **human / AI** classifier — "
                    "[GigaCheck-Classifier-Multi]"
                    "(https://huggingface.co/iitolstykh/GigaCheck-Classifier-Multi)"
                )
                build_classifier_tab()
            with gr.Tab("Detector"):
                gr.Markdown(
                    "Fine-grained **AI-span detector** — "
                    "[GigaCheck-Detector-Multi]"
                    "(https://huggingface.co/iitolstykh/GigaCheck-Detector-Multi)"
                )
                build_detector_tab()
    return demo


if __name__ == "__main__":
    build_demo().queue(max_size=50).launch()