Mundo oaxaca

# ─── Clasificador de Positividad Tóxica ───────────────────────
# Fine-tuning de RoBERTa para detección de invalidación emocional
# Requiere: transformers, datasets, torch

from transformers import AutoTokenizer, AutoModelForSequenceClassification
from torch.nn.functional import softmax
import torch

# Etiquetas del corpus anotado:
# 0 = psicología positiva legítima (valida emociones)
# 1 = positividad tóxica (invalida / culpabiliza)

MODEL_ID = "xlm-roberta-base"  # Multilingüe (ES, EN, PT, FR...)

tokenizer = AutoTokenizer.from_pretrained(MODEL_ID)
model     = AutoModelForSequenceClassification.from_pretrained(
    "./checkpoints/toxic_positivity_v2",  # Fine-tuned
    num_labels=2
)
model.eval()

def score_content(text: str) -> dict:
    """
    Retorna:
      toxic_score  (0–1): probabilidad de positividad tóxica
      label        : 'TOXIC_POSITIVE' | 'LEGITIMATE_POSITIVE'
      confidence   : confianza del modelo
    """
    inputs = tokenizer(text, return_tensors="pt",
                       truncation=True, max_length=256)
    with torch.no_grad():
        logits = model(**inputs).logits
    probs      = softmax(logits, dim=-1)[0]
    toxic_prob = probs[1].item()

    return {
        "toxic_score" : round(toxic_prob, 4),
        "label"       : "TOXIC_POSITIVE" if toxic_prob >= 0.65 else "LEGITIMATE_POSITIVE",
        "confidence"  : round(probs[int(toxic_prob >= 0.65)].item(), 4)
    }

# ─── Ejemplos ────────────────────────────────────────────────
print(score_content("Está bien sentirse triste; tus emociones son válidas."))
# → {'toxic_score': 0.08, 'label': 'LEGITIMATE_POSITIVE', 'confidence': 0.92}

print(score_content("¡Deja de quejarte! Las vibras negativas atraen fracaso."))
# → {'toxic_score': 0.94, 'label': 'TOXIC_POSITIVE', 'confidence': 0.94}

# ─── Emotional Harm Propagation Score ────────────────────────
# Combina toxicidad del contenido + alcance + vulnerabilidad
# del público para calcular el riesgo real de daño emocional

from dataclasses import dataclass
import math

@dataclass
class ContentMetrics:
    toxic_score      : float   # 0–1  (clasificador anterior)
    reach_estimate   : int     # usuarios únicos alcanzados
    minor_audience_pct: float  # % audiencia <18 años (0–1)
    velocity         : float   # shares/hora (viralidad)
    account_credibility: float # 0–1 (verified, historia, credenciales)

def compute_ehps(m: ContentMetrics) -> float:
    """
    EHPS = toxic_score
           × log10(reach + 1)          # escala logarítmica de alcance
           × (1 + minor_audience_pct)  # amplificación por audiencia vulnerable
           × velocity_factor            # urgencia de intervención
           × (1 - account_credibility) # descuento por credibilidad real
    
    Rango resultante: 0 (sin riesgo) → ∞ (intervención urgente)
    Umbral operativo recomendado: EHPS ≥ 3.5 → intervención automática
    """
    velocity_factor = 1 + math.log1p(m.velocity / 10)

    ehps = (
        m.toxic_score
        * math.log10(m.reach_estimate + 1)
        * (1 + m.minor_audience_pct)
        * velocity_factor
        * (1 - m.account_credibility)
    )
    return round(ehps, 3)

def intervention_level(ehps: float) -> str:
    if   ehps < 1.0: return "NONE       — Sin intervención"
    elif ehps < 2.0: return "LEVEL_1    — Etiqueta informativa"
    elif ehps < 3.5: return "LEVEL_2    — Reducción de alcance algorítmico"
    elif ehps < 5.0: return "LEVEL_3    — Recurso de salud mental adjunto"
    else          : return "LEVEL_4    — Revisión humana urgente"

# ─── Caso simulado: cuenta MLM, 800k seguidores, 40% menores ──
metrics = ContentMetrics(
    toxic_score=0.91,
    reach_estimate=820_000,
    minor_audience_pct=0.40,
    velocity=3400,   # shares/hora
    account_credibility=0.05
)
score = compute_ehps(metrics)
print(f"EHPS: {score} → {intervention_level(score)}")
# → EHPS: 7.842 → LEVEL_4 — Revisión humana urgente

# ─── Perspective API (Google/Jigsaw) ─────────────────────────
# Detecta atributos de toxicidad relacional complementarios
# Documentación: https://developers.perspectiveapi.com/

import requests, os

PERSPECTIVE_KEY = os.environ["PERSPECTIVE_API_KEY"]
ENDPOINT = "https://commentanalyzer.googleapis.com/v1alpha1/comments:analyze"

def analyze_with_perspective(text: str, lang: str = "es") -> dict:
    """
    Atributos relevantes para positividad tóxica:
    - TOXICITY             : toxicidad general
    - INSULT               : invalidación/menosprecio emocional
    - IDENTITY_ATTACK      : culpabilización por identidad (ej. "los pobres no trabajan")
    - INFLAMMATORY         : provocación emocional
    """
    payload = {
        "comment": {"text": text},
        "languages": [lang],
        "requestedAttributes": {
            "TOXICITY"        : {},
            "INSULT"          : {},
            "IDENTITY_ATTACK" : {},
            "INFLAMMATORY"    : {}
        }
    }
    r = requests.post(
        ENDPOINT, params={"key": PERSPECTIVE_KEY}, json=payload
    )
    scores = r.json()["attributeScores"]
    return {
        attr: round(data["summaryScore"]["value"], 3)
        for attr, data in scores.items()
    }

# ─── Ejemplo ──────────────────────────────────────────────────
result = analyze_with_perspective(
    "Si no eres exitoso es porque no tienes suficiente mentalidad ganadora."
)
print(result)
# {
#   'TOXICITY': 0.78, 'INSULT': 0.71,
#   'IDENTITY_ATTACK': 0.65, 'INFLAMMATORY': 0.82
# }

# ─── Pipeline Completo de Intervención ───────────────────────
# Orquesta todos los módulos y toma decisiones de intervención

from toxic_positivity_classifier import score_content
from ehps_calculator import compute_ehps, intervention_level, ContentMetrics
from perspective_integration import analyze_with_perspective
from platform_adapters import TikTokAdapter, MetaAdapter, YouTubeAdapter

MENTAL_HEALTH_RESOURCES = {
    "es": "https://www.crisistext.org | Línea de Bienestar: 800-290-0024",
    "en": "https://www.crisistextline.org | Text HOME to 741741",
    "pt": "https://www.cvv.org.br | CVV: 188"
}

class ToxicPositivityPipeline:

    def __init__(self, platform: str):
        self.platform = platform
        self.adapter  = {
            "tiktok"   : TikTokAdapter,
            "meta"     : MetaAdapter,
            "youtube"  : YouTubeAdapter
        }[platform]()

    def process(self, content_id: str) -> dict:
        # 1. Extracción
        raw        = self.adapter.fetch(content_id)
        text       = raw["text"] + " " + raw.get("transcript", "")
        lang       = raw.get("lang", "es")

        # 2. Scoring
        tp_result  = score_content(text)
        pv_result  = analyze_with_perspective(text, lang)

        # 3. Score combinado (promedio ponderado)
        combined_toxic = (
            tp_result["toxic_score"] * 0.5 +
            pv_result["TOXICITY"]     * 0.25 +
            pv_result["INSULT"]        * 0.15 +
            pv_result["INFLAMMATORY"]  * 0.10
        )

        # 4. EHPS
        metrics = ContentMetrics(
            toxic_score=combined_toxic,
            reach_estimate=raw["reach"],
            minor_audience_pct=raw["minor_pct"],
            velocity=raw["shares_per_hour"],
            account_credibility=raw["credibility_score"]
        )
        ehps  = compute_ehps(metrics)
        level = intervention_level(ehps)

        # 5. Intervención
        if   ehps >= 5.0: self.adapter.flag_for_review(content_id, ehps)
        elif ehps >= 3.5: self.adapter.attach_resource(
                             content_id, MENTAL_HEALTH_RESOURCES[lang])
        elif ehps >= 2.0: self.adapter.reduce_reach(content_id, factor=0.4)
        elif ehps >= 1.0: self.adapter.add_label(
                             content_id, "Este contenido puede no reflejar experiencias diversas")

        return {"content_id": content_id, "ehps": ehps,
                "level": level, "platform": self.platform}