""" src/app/streamlit_app.py App SignalMod — detección de hate speech estilo YouTube. Ejecutar: streamlit run src/app/streamlit_app.py """ import html import sys import random import datetime from pathlib import Path import streamlit as st import pandas as pd from transformers.utils import logging logging.set_verbosity_error() # ── Paths ───────────────────────────────────────────────────────────────────── PROJECT_ROOT = Path(__file__).resolve().parents[2] sys.path.insert(0, str(PROJECT_ROOT)) try: from src.service.model_service import ModelService, AVAILABLE_MODELS except ImportError: sys.path.insert(0, str(Path(__file__).parent.parent)) from service.model_service import ModelService, AVAILABLE_MODELS # ── Config ──────────────────────────────────────────────────────────────────── st.set_page_config( page_title="SignalMod", page_icon="🎬", layout="wide", initial_sidebar_state="expanded", ) # ── CSS ─────────────────────────────────────────────────────────────────────── # Nota: NO ocultamos el header completo para preservar el botón de toggle del sidebar. # Solo ocultamos el menú hamburguesa y el footer de Streamlit. st.markdown(""" """, unsafe_allow_html=True) # ── Session state init ──────────────────────────────────────────────────────── def _init_state(): defaults = { "page" : "Home", "selected_model": list(AVAILABLE_MODELS.keys())[0], "threshold" : 0.5, "pending_modal" : None, # dict con el comentario pendiente de decisión "comments": [ {"user": "user_prime", "initial": "U", "text": "Excelente video, muy informativo!", "time": "1 h", "is_toxic": False, "probability": 0.04, "labels": []}, {"user": "troll_master", "initial": "T", "text": "Esto es una basura completa", "time": "30 min", "is_toxic": True, "probability": 0.91, "labels": ["Insulto","Agresividad"]}, {"user": "curious_viewer", "initial": "C", "text": "¿Alguien puede explicar esto mejor?", "time": "15 min", "is_toxic": False, "probability": 0.07, "labels": []}, ], "hub_history": [ {"Usuario": "@user_992", "Comentario": '"No puedo creer que seas tan..."', "Score": 0.94, "Acción": "🚫 Bloqueado"}, {"Usuario": "@alpha_mod", "Comentario": '"Spam repetitivo de enlaces."', "Score": 0.82, "Acción": "🚩 Revisión"}, {"Usuario": "@anon_404", "Comentario": '"Discurso de odio en contexto."', "Score": 0.98, "Acción": "📋 Archivado"}, {"Usuario": "@user_123", "Comentario": '"¡Gran contenido, sigan!"', "Score": 0.03, "Acción": "✅ Aprobado"}, {"Usuario": "@viewer_x", "Comentario": '"Esta gente debería desaparecer."',"Score": 0.97, "Acción": "🚫 Bloqueado"}, ], } for k, v in defaults.items(): if k not in st.session_state: st.session_state[k] = v _init_state() # ── Model cache ─────────────────────────────────────────────────────────────── @st.cache_resource(show_spinner="Cargando modelo...") def get_service(model_name: str) -> ModelService: return ModelService(model_name, PROJECT_ROOT) # ══════════════════════════════════════════════════════════════════════════════ # SIDEBAR # ══════════════════════════════════════════════════════════════════════════════ def render_sidebar(): with st.sidebar: # Logo st.markdown( "
" "🎬 SignalMod
" "
" "Signal within the Noise
", unsafe_allow_html=True, ) nav = {"Home": "🏠", "Moderator Hub": "📊", "Settings": "⚙️"} for page, icon in nav.items(): label = f"{icon} {page}" clicked = st.button(label, key=f"nav_{page}", use_container_width=True) if clicked: st.session_state.page = page st.rerun() st.divider() # Info modelo activo model_short = st.session_state.selected_model.split("(")[0].strip() tox_cnt = sum(1 for c in st.session_state.comments if c["is_toxic"]) total_c = len(st.session_state.comments) st.markdown( f"", unsafe_allow_html=True, ) # ══════════════════════════════════════════════════════════════════════════════ # MODAL — toxicidad detectada # ══════════════════════════════════════════════════════════════════════════════ @st.dialog("⚠️ Aviso de Toxicidad Detectada") def show_toxicity_modal(): """ @st.dialog crea una ventana modal nativa de Streamlit (1.32+). Cuando se llama a la función decorada, Streamlit renderiza el contenido dentro de un overlay modal y pausa la ejecución normal del script. """ data = st.session_state.pending_modal if not data: st.rerun() return text = data["text"] prob = data["probability"] lbls = data["labels"] pct = int(prob * 100) color = "#cc0000" if pct >= 70 else "#ff6d00" if pct >= 40 else "#f5a623" st.markdown( "
⚠️
", unsafe_allow_html=True, ) st.markdown( f"
" f""{html.escape(text[:140])}{'...' if len(text)>140 else ''}"
", unsafe_allow_html=True, ) # Barra de toxicidad st.markdown( f"
" f"ÍNDICE DE TOXICIDAD" f"{pct}%
" f"
" f"
" f"
", unsafe_allow_html=True, ) # Etiquetas if lbls: tags = " ".join( f"" f"🚩 {html.escape(l)}" for l in lbls ) st.markdown(f"
{tags}
", unsafe_allow_html=True) st.markdown("
", unsafe_allow_html=True) col1, col2 = st.columns(2) with col1: if st.button("✏️ Editar comentario", use_container_width=True, type="primary"): st.session_state.pending_modal = None st.rerun() with col2: if st.button("Publicar de todas maneras", use_container_width=True): # Publicar aunque sea tóxico c = st.session_state.pending_modal st.session_state.comments.append(c) st.session_state.hub_history.insert(0, { "Usuario" : "@usuario", "Comentario": f'"{c["text"][:45]}..."', "Score" : round(c["probability"], 2), "Acción" : "⚠️ Override usuario", }) st.session_state.pending_modal = None st.rerun() # ══════════════════════════════════════════════════════════════════════════════ # HOME — interfaz estilo YouTube # ══════════════════════════════════════════════════════════════════════════════ def render_home(): # Disparar modal si hay comentario pendiente if st.session_state.pending_modal: show_toxicity_modal() col_main, col_right = st.columns([2.8, 1], gap="large") with col_main: # Video st.markdown( "
▶
", unsafe_allow_html=True, ) st.markdown( "
AI Moderation Demo — Detección de Hate Speech en tiempo real
" "
15k vistas · 2 horas atrás
", unsafe_allow_html=True, ) row_ch, row_sub = st.columns([3, 1]) with row_ch: st.markdown( "
" "
S
" "
SignalMod AI
" "
1.2M suscriptores
", unsafe_allow_html=True, ) st.divider() # ── Comentarios ──────────────────────────────────────────────────── tox_cnt = sum(1 for c in st.session_state.comments if c["is_toxic"]) st.markdown( f"
{len(st.session_state.comments)} Comentarios " f"· {tox_cnt} detectados
", unsafe_allow_html=True, ) # Input de nuevo comentario new_text = st.text_area( "Escribe un comentario...", height=80, label_visibility="collapsed", key="comment_input", placeholder="Escribe un comentario...", ) # Análisis en tiempo real (solo cuando hay texto) analysis = None if new_text.strip(): svc = get_service(st.session_state.selected_model) analysis = svc.predict(new_text) pct = int(analysis["probability"] * 100) color = "#cc0000" if pct >= 70 else "#f5a623" if pct >= 40 else "#00c853" verdict = "TÓXICO" if analysis["is_toxic"] else "SEGURO" v_color = "#cc0000" if analysis["is_toxic"] else "#00c853" st.markdown( f"
" f"🔍 Analizando..." f"{verdict}" f"Toxicidad: {pct}%" f"
" f"
" f"
", unsafe_allow_html=True, ) col_c, col_p = st.columns([1, 1]) with col_c: if st.button("Cancelar", use_container_width=True): st.rerun() with col_p: post = st.button("Comentar", type="primary", use_container_width=True) # Procesar envío if post and new_text.strip(): if analysis is None: svc = get_service(st.session_state.selected_model) analysis = svc.predict(new_text) comment_obj = { "user" : "usuario", "initial" : "U", "text" : new_text.strip(), "time" : "ahora", "is_toxic" : analysis["is_toxic"], "probability": analysis["probability"], "labels" : analysis["labels"], } if analysis["is_toxic"]: # Guardar en pendiente y mostrar modal en el próximo render st.session_state.pending_modal = comment_obj st.rerun() else: # Publicar directamente st.session_state.comments.append(comment_obj) st.session_state.hub_history.insert(0, { "Usuario" : "@usuario", "Comentario": f'"{new_text.strip()[:45]}{"..." if len(new_text)>45 else ""}"', "Score" : round(analysis["probability"], 2), "Acción" : "✅ Aprobado", }) st.rerun() # ── Lista de comentarios ─────────────────────────────────────────── for c in reversed(st.session_state.comments): is_tox = c["is_toxic"] pct = int(c["probability"] * 100) av_class = "c-avatar" if is_tox else "c-avatar safe" badge = ( "TÓXICO" if is_tox else "SEGURO" ) text_class = "c-text toxic" if is_tox else "c-text" flagged = "
🚩 Flagged for review
" if is_tox else "" # html.escape() protege contra caracteres que rompen el HTML safe_text = html.escape(c["text"]) safe_user = html.escape(c["user"]) initial = html.escape(c.get("initial", c["user"][0].upper())) st.markdown( f"
" f"
{initial}
" f"
" f"
" f" @{safe_user}" f" {c['time']}" f" {badge}" f"
" f"
{safe_text}
" f" {flagged}" f"
" f"
", unsafe_allow_html=True, ) # ── Columna derecha ──────────────────────────────────────────────────── with col_right: st.markdown("**Sugeridos**") suggested = [ ("🤖", "Understanding Transformer Models...", "Neural Systems", "89k · 1 día"), ("🎓", "The Future of Content Moderation", "Tech Ethics Pro", "1.4M · 2 sem"), ("📡", "Signal vs Noise: SignalMod Deep Dive","SignalMod AI", "250k · 3 días"), ("💡", "Why AI Moderation is Harder Than...", "Ethics in Code", "45k · 5 h"), ("🔬", "Hate Speech Detection 2024", "AI Research Lab", "12k · 1 sem"), ] for emoji, title, ch, meta in suggested: st.markdown( f"
" f"
{emoji}
" f"
" f"
{html.escape(title)}
" f"
{html.escape(ch)}
" f"
{html.escape(meta)}
" f"
" f"
", unsafe_allow_html=True, ) # ══════════════════════════════════════════════════════════════════════════════ # MODERATOR HUB # ══════════════════════════════════════════════════════════════════════════════ def render_hub(): try: import plotly.graph_objects as go except ImportError: st.error("Instala plotly: pip install plotly") return st.markdown("## 📊 Panel de Estadísticas") # ── Cards de configuración ────────────────────────────────────────────── model_short = st.session_state.selected_model.split("(")[0].strip() c1, c2, c3 = st.columns(3) for col, label, val in [ (c1, "MODEL ARCHITECTURE", model_short), (c2, "CONFIDENCE THRESHOLD", f"{st.session_state.threshold:.2f} Alpha"), (c3, "LANGUAGE COVERAGE", "English"), ]: with col: st.markdown( f"
" f"
{label}
" f"
" f"{html.escape(str(val))}
", unsafe_allow_html=True, ) st.write("") # ── KPIs ─────────────────────────────────────────────────────────────── total = len(st.session_state.comments) + 100 tox_cnt = sum(1 for c in st.session_state.comments if c["is_toxic"]) + 5 tox_rate = tox_cnt / total * 100 m1, m2, m3 = st.columns(3) m1.metric("💬 Total comentarios", f"{total:,}", "+12%") m2.metric("☠️ Tasa de toxicidad", f"{tox_rate:.1f}%", f"+0.8%", delta_color="inverse") m3.metric("🎯 F1 Score", "0.7579", "Stable") st.divider() # ── Gráficos ─────────────────────────────────────────────────────────── gcol, pcol = st.columns([2.2, 1]) with gcol: days = ["Lun","Mar","Mié","Jue","Vie","Sáb","Dom"] vals = [random.randint(30, 80) for _ in days] vals[3] = max(vals) + 25 colors = ["#cc0000" if i == 3 else "#b3c6ff" for i in range(7)] fig = go.Figure(go.Bar(x=days, y=vals, marker_color=colors, width=0.55)) fig.update_layout( title="Tendencias de Toxicidad (7D)", paper_bgcolor="#ffffff", plot_bgcolor="#ffffff", margin=dict(l=20, r=20, t=40, b=20), height=260, font=dict(size=11, color="#0f0f0f"), ) fig.update_yaxes(showgrid=True, gridcolor="#f0f0f0", zeroline=False) fig.update_xaxes(showgrid=False) st.plotly_chart(fig, use_container_width=True) with pcol: fig2 = go.Figure(go.Pie( labels=["Hate Speech","Insulto","Agresividad"], values=[45, 35, 20], hole=0.58, marker_colors=["#cc0000","#0f0f0f","#909090"], textfont_size=11, )) fig2.update_layout( title="Categorías", paper_bgcolor="#ffffff", margin=dict(l=10, r=10, t=40, b=10), height=260, legend=dict(font=dict(size=10), orientation="v"), font=dict(size=11, color="#0f0f0f"), ) st.plotly_chart(fig2, use_container_width=True) # ── Historial ────────────────────────────────────────────────────────── st.markdown("### Historial Reciente") df = pd.DataFrame(st.session_state.hub_history) if not df.empty: st.dataframe( df, use_container_width=True, hide_index=True, column_config={ "Score": st.column_config.ProgressColumn( "Score", min_value=0, max_value=1, format="%.2f" ) }, ) # ══════════════════════════════════════════════════════════════════════════════ # SETTINGS # ══════════════════════════════════════════════════════════════════════════════ def render_settings(): st.markdown("## ⚙️ Ajustes") # ── Selección de modelo ───────────────────────────────────────────────── st.markdown("### 🤖 Modelo de detección",) st.caption( "Los modelos HuggingFace se descargan la primera vez (~300–600 MB). " "Requieren: `pip install transformers torch sentencepiece`" ) st.write("") # Usamos st.radio para la selección — sin bugs de HTML model_names = list(AVAILABLE_MODELS.keys()) current_idx = model_names.index(st.session_state.selected_model) \ if st.session_state.selected_model in model_names else 0 chosen = st.radio( "Seleccionar modelo", model_names, index=current_idx, label_visibility="collapsed", ) if chosen != st.session_state.selected_model: st.session_state.selected_model = chosen st.rerun() # Ficha del modelo seleccionado info = AVAILABLE_MODELS[st.session_state.selected_model] st.markdown( f"
" f"
{info['icon']} {html.escape(st.session_state.selected_model)}
" f"
{html.escape(info['description'])}
" f"
" f"⚡ {html.escape(info['speed'])}" f"🎯 {html.escape(info['accuracy'])}" f"📦 {html.escape(info['requires'])}" f"
", unsafe_allow_html=True, ) # Info sobre modelo fine-tuneado if st.session_state.selected_model == "Modelo fine-tuneado (local)": path = PROJECT_ROOT / "models" / "finetuned_hf" if path.exists(): st.success(f"✅ Modelo encontrado en `{path}`") else: st.warning( f"⚠️ No se encontró el modelo en `{path}`. " f"Ejecuta el **notebook 08** para generar el modelo fine-tuneado." ) st.divider() # ── Umbral de confianza ───────────────────────────────────────────────── st.markdown("### 🎚️ Umbral de confianza") st.caption("Probabilidad mínima para marcar un comentario como tóxico.") new_thr = st.slider( "Umbral", min_value=0.3, max_value=0.9, step=0.05, value=st.session_state.threshold, label_visibility="collapsed", format="%.2f", ) if new_thr != st.session_state.threshold: st.session_state.threshold = new_thr st.info(f"Umbral actualizado: **{new_thr:.2f}**") ta, tb = st.columns(2) ta.info(f"⬇️ **{new_thr:.2f}** bajo → más FP (más censura)", icon="⚠️") tb.info(f"⬆️ **{new_thr:.2f}** alto → más FN (más escapes)", icon="⚠️") st.divider() # ── Test rápido ───────────────────────────────────────────────────────── st.markdown("### 🧪 Probar modelo") test_txt = st.text_input( "Texto a analizar", placeholder="Ej: This is absolutely stupid and racist...", label_visibility="collapsed", ) if st.button("Analizar", type="primary") and test_txt.strip(): with st.spinner("Analizando..."): svc = get_service(st.session_state.selected_model) res = svc.predict(test_txt) pct = int(res["probability"] * 100) verdict = "🔴 TÓXICO" if res["is_toxic"] else "🟢 SEGURO" st.markdown(f"**{verdict}** — {pct}% de toxicidad") st.progress(res["probability"]) if res["labels"]: st.markdown(f"**Categorías:** {', '.join(res['labels'])}") if "error" in res: st.error(f"Error: {res['error']}") st.caption(f"Modelo: {res['model_used']}") # ══════════════════════════════════════════════════════════════════════════════ # MAIN # ══════════════════════════════════════════════════════════════════════════════ def main(): render_sidebar() page = st.session_state.page if page == "Home": render_home() elif page == "Moderator Hub": render_hub() elif page == "Settings": render_settings() if __name__ == "__main__": main()