app.py

import torch
import torch.nn.functional as F
from transformers import AutoTokenizer, AutoModelForSequenceClassification
import re
import pandas as pd
import gradio as gr

# -----------------------------
# MODEL (Fakespot 2025)
# -----------------------------
MODEL_NAME = "fakespot-ai/roberta-base-ai-text-detection-v1"
tokenizer = AutoTokenizer.from_pretrained(MODEL_NAME)

device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
dtype = torch.bfloat16 if (device.type == "cuda" and torch.cuda.is_bf16_supported()) else torch.float32

model = AutoModelForSequenceClassification.from_pretrained(
    MODEL_NAME, dtype=dtype
).to(device).eval()

# -----------------------------
# AI DECISION THRESHOLD (80%)
# -----------------------------
THRESHOLD = 0.80

# -----------------------------
# SENTENCE SPLITTING UTILITIES
# -----------------------------
ABBR = [
    "e.g", "i.e", "mr", "mrs", "ms", "dr", "prof", "vs", "etc", "fig", "al",
    "jr", "sr", "st", "no", "vol", "pp", "mt", "inc", "ltd", "co",
    "u.s", "u.k", "a.m", "p.m"
]
ABBR_REGEX = re.compile(
    r"\b(" + "|".join(map(re.escape, ABBR)) + r")\.",
    flags=re.IGNORECASE
)

def _protect(text: str) -> str:
    t = text.strip()
    if not t:
        return ""
    t = re.sub(r"\s*\n+\s*", " ", t)
    t = t.replace("...", "⟨ELLIPSIS⟩")
    t = re.sub(r"(?<=\d)\.(?=\d)", "⟨DECIMAL⟩", t)
    t = ABBR_REGEX.sub(r"\1⟨ABBRDOT⟩", t)
    return t

def _restore(text: str) -> str:
    return (
        text.replace("⟨ABBRDOT⟩", ".")
            .replace("⟨DECIMAL⟩", ".")
            .replace("⟨ELLIPSIS⟩", "...")
    )

def sentence_split(text: str):
    t = _protect(text)
    if not t:
        return []

    # hard sentence boundary detection
    parts = re.split(r"([.?!])\s+(?=[\"“”‘’']?\s*[A-Z(]|$)", t)

    sentences, buf = [], ""
    for i, chunk in enumerate(parts):
        if i % 2 == 0:
            buf += chunk
        else:
            buf += chunk
            sentences.append(buf.strip())
            buf = ""

    if buf.strip():
        sentences.append(buf.strip())

    return [_restore(s).strip() for s in sentences if s.strip()]


# -----------------------------
# CORE ANALYSIS — PER SENTENCE
# -----------------------------
def analyze(text, max_len=512):
    sents = sentence_split(text)
    if not sents:
        return "—", "—", "<em>Paste some text to analyze.</em>", None

    clean_sents = [re.sub(r"\s+", " ", s).strip() for s in sents]

    # tokenize list of sentences
    inputs = tokenizer(
        clean_sents,
        return_tensors="pt",
        padding=True,
        truncation=True,
        max_length=max_len
    ).to(device)

    # model inference (per sentence)
    with torch.no_grad():
        logits = model(**inputs).logits
        probs = F.softmax(logits, dim=-1)[:, 1].detach().cpu().tolist()

    # overall AI score
    overall_ai = sum(probs) / len(probs)
    overall_pct = f"{overall_ai * 100:.1f}%"

    overall_label = (
        "🤖 Likely AI Written" if overall_ai >= THRESHOLD else "🧒 Likely Human Written"
    )

    # highlights + table
    rows, highlights = [], []

    for i, sentence in enumerate(sents, start=1):
        ai_p = float(probs[i - 1])
        pct = f"{ai_p * 100:.1f}%"

        label = "AI" if ai_p >= THRESHOLD else "Human"

        # colors
        if ai_p < 0.30:
            color = "#11823b"
        elif ai_p < 0.70:
            color = "#b8860b"
        else:
            color = "#b80d0d"

        normalized = re.sub(r"\s+", " ", sentence)

        highlights.append(
            "<div style='margin:6px 0; padding:6px 8px; border-radius:6px;"
            "background:rgba(0,0,0,0.03)'>"
            f"<strong style='color:{color}'>[{pct} {label}]</strong> "
            f"{normalized}</div>"
        )

        rows.append([i, sentence, round(ai_p, 4), label])

    df = pd.DataFrame(rows, columns=["#", "Sentence", "AI_Prob", "Label"])
    html = "\n".join(highlights)

    return overall_label, overall_pct, html, df


# -----------------------------
# GRADIO UI
# -----------------------------
with gr.Blocks() as demo:
    gr.Markdown("### 🕵️ AI Written Text Detector — Fakespot Model (80% Threshold)")

    text_input = gr.Textbox(label="Paste text", lines=14, placeholder="Your content…")
    btn = gr.Button("Analyze")

    verdict = gr.Label(label="Verdict (Overall)")
    score = gr.Label(label="AI Score (Average across sentences)")
    highlights = gr.HTML(label="Per-Sentence Highlights")
    table = gr.Dataframe(headers=["#", "Sentence", "AI_Prob", "Label"], wrap=True)

    btn.click(analyze, inputs=[text_input], outputs=[verdict, score, highlights, table])

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