Spaces:
Sleeping
Sleeping
Download app.py from Danielamn026/crisisbench-classifier: direct link, hf CLI and curl.
- Browser
- Download file 12.5 kB
-
https://huggingface.co/spaces/Danielamn026/crisisbench-classifier/resolve/main/app.py
- Command line
-
hf download hf://spaces/Danielamn026/crisisbench-classifier/app.py
-
curl -L -o app.py https://huggingface.co/spaces/Danielamn026/crisisbench-classifier/resolve/main/app.py
12.5 kB
| # Clasificador de mensajes en emergencias - CrisisBench | |
| # Proyecto final Tecnicas de Aprendizaje de Maquina - PUJ 2026-2 | |
| import re | |
| import html | |
| import ftfy | |
| import pickle | |
| import torch | |
| import gradio as gr | |
| import spaces | |
| from transformers import AutoTokenizer, AutoModelForSequenceClassification | |
| # ============================================================ | |
| # 1. PREPROCESAMIENTO | |
| # ============================================================ | |
| RT_RE = re.compile(r'^(rt\s+@\w+:?\s*)+') | |
| URL_RE = re.compile(r'https?://\S+|www\.\S+') | |
| USER_RE = re.compile(r'@\w+') | |
| HASH_RE = re.compile(r'#(\w+)') | |
| WS_RE = re.compile(r'\s+') | |
| def limpiar(t: str) -> str: | |
| t = ftfy.fix_text(t) | |
| t = html.unescape(t) | |
| t = RT_RE.sub("", t) | |
| t = URL_RE.sub(" <URL> ", t) | |
| t = USER_RE.sub(" <USER> ", t) | |
| t = HASH_RE.sub(lambda m: " " + re.sub(r"(?<=[a-z])(?=[A-Z])", " ", m.group(1)) + " ", t) | |
| return WS_RE.sub(" ", t).strip() | |
| # ============================================================ | |
| # 2. CARGAR MODELOS | |
| # ============================================================ | |
| print("Cargando XLM-R...") | |
| XLM_PATH = "Danielamn026/crisisbench-xlmr" | |
| tok = AutoTokenizer.from_pretrained(XLM_PATH) | |
| model_xlm = AutoModelForSequenceClassification.from_pretrained(XLM_PATH) | |
| model_xlm.eval() | |
| print("XLM-R listo") | |
| print("Cargando LogReg + TF-IDF...") | |
| with open("logreg.pkl", "rb") as f: | |
| logreg = pickle.load(f) | |
| with open("tfidf_vectorizer.pkl", "rb") as f: | |
| vec = pickle.load(f) | |
| print("LogReg listo") | |
| MAX_LEN = 64 | |
| # ============================================================ | |
| # 3. HEURISTICA: detectar mensajes propensos a etiqueta dudosa | |
| # ============================================================ | |
| # Patrones observados en la revision manual de 100 mensajes (seccion 4.8 del cuaderno) | |
| # donde el dataset marca "informativo" pero son condolencias/reacciones emocionales | |
| PATRONES_CONDOLENCIA = [ | |
| r'\b(pray|prayers|praying|rezar|rezando|oraciones?)\b', | |
| r'\b(fuerza|animo|anim[oó])\b', | |
| r'\b(thoughts?|pensamientos?)\s+(and|y)\s+(prayers?|oraciones?)\b', | |
| r'\b(condolencias?|condolences?)\b', | |
| r'\b(tristeza|triste|sad|sorry|lo\s+siento)\b', | |
| r'\b(que\s+(horror|tragedia|pena|lastima|desgracia))\b', | |
| r'\b(rest\s+in\s+peace|rip|descanse(n)?\s+en\s+paz|dep)\b', | |
| r'#(pray|fuerza|prayfor|rip\w*|descansa)', | |
| ] | |
| RE_CONDOLENCIA = re.compile("|".join(PATRONES_CONDOLENCIA), re.IGNORECASE) | |
| # Palabras tipicas de mensajes genuinamente operativos (si aparecen, es menos probable que sea solo condolencia) | |
| PATRONES_OPERATIVO = [ | |
| r'\b\d+\s*(muertos?|heridos?|desaparecidos?|dead|injured|missing|casualties?|victims?)\b', | |
| r'\b(evacuar|evacuation|rescate|rescue|emergencia|emergency|911|112|hospital)\b', | |
| r'\b(magnitud|magnitude|epicentro|epicenter|richter|escala)\b', | |
| r'\b(colapso|collapsed|destruyo|destroyed|derrumbe|hundio|sank)\b', | |
| r'\b(alert|alerta|aviso|warning|tsunami|replicas?|aftershock)\b', | |
| ] | |
| RE_OPERATIVO = re.compile("|".join(PATRONES_OPERATIVO), re.IGNORECASE) | |
| def detectar_caso_dudoso(texto: str, p_xlm: float, p_lr: float) -> str: | |
| """Devuelve HTML de aviso si el caso puede ser una etiqueta dudosa del dataset.""" | |
| tiene_condolencia = bool(RE_CONDOLENCIA.search(texto)) | |
| tiene_operativo = bool(RE_OPERATIVO.search(texto)) | |
| ambos_altos = p_xlm >= 0.5 and p_lr >= 0.5 | |
| # Caso tipico: parece condolencia, NO tiene datos operativos, y los modelos lo marcan como informativo | |
| if tiene_condolencia and not tiene_operativo and ambos_altos: | |
| return """ | |
| <div style="padding: 14px 18px; background: rgba(239, 68, 68, 0.08); border-left: 3px solid #ef4444; border-radius: 8px; color: #fca5a5; font-size: 13.5px; margin-top: 12px; line-height: 1.5;"> | |
| <strong>⚠ Atencion: este es probablemente un caso de etiqueta dudosa del dataset.</strong><br> | |
| El mensaje contiene expresiones de condolencia o reaccion emocional (sin datos operativos concretos). | |
| La revision manual mostro que el 32% de los mensajes en espanol marcados como "informativo" en CrisisBench | |
| son en realidad de este tipo. Los modelos aprendieron ese sesgo. <strong>Para uso real, un humano deberia descartarlo.</strong> | |
| </div> | |
| """ | |
| return "" | |
| # ============================================================ | |
| # 4. RENDER DE RESULTADOS | |
| # ============================================================ | |
| def render_resultado(p_inf: float) -> str: | |
| es_inf = p_inf >= 0.5 | |
| pct = p_inf * 100 | |
| chip_bg = "#10b981" if es_inf else "#6b7280" | |
| chip_label = "INFORMATIVO" if es_inf else "NO INFORMATIVO" | |
| chip_icon = "✓" if es_inf else "✕" | |
| return f""" | |
| <div style="padding: 16px; border-radius: 12px; background: rgba(255,255,255,0.03); border: 1px solid rgba(255,255,255,0.08);"> | |
| <div style="display: inline-block; padding: 6px 14px; background: {chip_bg}; color: white; border-radius: 999px; font-weight: 700; font-size: 13px; letter-spacing: 0.5px;"> | |
| {chip_icon} {chip_label} | |
| </div> | |
| <div style="margin-top: 14px; font-size: 13px; color: #9ca3af;">Probabilidad de ser informativo</div> | |
| <div style="display: flex; align-items: center; gap: 12px; margin-top: 6px;"> | |
| <div style="flex: 1; background: rgba(255,255,255,0.08); height: 10px; border-radius: 999px; overflow: hidden;"> | |
| <div style="width: {pct:.1f}%; height: 100%; background: linear-gradient(90deg, {chip_bg}dd, {chip_bg}); border-radius: 999px;"></div> | |
| </div> | |
| <div style="font-size: 20px; font-weight: 700; color: white; min-width: 70px; text-align: right;">{pct:.1f}%</div> | |
| </div> | |
| </div> | |
| """ | |
| def render_acuerdo(pred_xlm_inf, pred_lr_inf) -> str: | |
| if pred_xlm_inf == pred_lr_inf: | |
| return """ | |
| <div style="padding: 12px 16px; background: rgba(16, 185, 129, 0.1); border-left: 3px solid #10b981; border-radius: 8px; color: #6ee7b7; font-size: 14px;"> | |
| ✓ <strong>Los dos modelos coinciden.</strong> | |
| </div> | |
| """ | |
| return """ | |
| <div style="padding: 12px 16px; background: rgba(245, 158, 11, 0.1); border-left: 3px solid #f59e0b; border-radius: 8px; color: #fcd34d; font-size: 14px;"> | |
| ⚠ <strong>Los modelos discrepan.</strong> Suele indicar un mensaje ambiguo o un vocabulario poco conocido para el modelo clasico. | |
| </div> | |
| """ | |
| # ============================================================ | |
| # 5. INFERENCIA | |
| # ============================================================ | |
| def clasificar(mensaje: str): | |
| if not mensaje or not mensaje.strip(): | |
| return "", "<div style='color:#9ca3af'>Esperando mensaje...</div>", "<div style='color:#9ca3af'>Esperando mensaje...</div>", "" | |
| texto_limpio = limpiar(mensaje) | |
| device = "cuda" if torch.cuda.is_available() else "cpu" | |
| model_xlm.to(device) | |
| with torch.no_grad(): | |
| enc = tok(texto_limpio, padding="max_length", truncation=True, | |
| max_length=MAX_LEN, return_tensors="pt").to(device) | |
| logits = model_xlm(**enc).logits | |
| probs_xlm = torch.softmax(logits, dim=-1).cpu().numpy()[0] | |
| p_xlm = float(probs_xlm[1]) | |
| X = vec.transform([texto_limpio]) | |
| p_lr = float(logreg.predict_proba(X)[0, 1]) | |
| aviso_html = render_acuerdo(p_xlm >= 0.5, p_lr >= 0.5) + detectar_caso_dudoso(mensaje, p_xlm, p_lr) | |
| return ( | |
| texto_limpio, | |
| render_resultado(p_xlm), | |
| render_resultado(p_lr), | |
| aviso_html, | |
| ) | |
| # ============================================================ | |
| # 6. INTERFAZ | |
| # ============================================================ | |
| CUSTOM_CSS = """ | |
| .gradio-container { font-family: 'Inter', -apple-system, BlinkMacSystemFont, sans-serif !important; } | |
| #header-banner { | |
| background: linear-gradient(135deg, #dc2626 0%, #f59e0b 100%); | |
| padding: 32px 28px; | |
| border-radius: 16px; | |
| margin-bottom: 24px; | |
| color: white; | |
| box-shadow: 0 10px 30px rgba(220, 38, 38, 0.25); | |
| } | |
| #header-banner h1 { margin: 0; font-size: 28px; font-weight: 800; color: white !important; } | |
| #header-banner p { margin: 8px 0 0 0; opacity: 0.95; font-size: 15px; color: white !important; } | |
| .model-title { font-size: 15px; font-weight: 700; margin-bottom: 4px; color: #f3f4f6; } | |
| .model-subtitle { font-size: 12px; color: #9ca3af; margin-bottom: 12px; } | |
| #footer-info { background: rgba(255,255,255,0.02); border-radius: 12px; padding: 20px 24px; margin-top: 24px; border: 1px solid rgba(255,255,255,0.05); } | |
| #footer-info h3 { margin-top: 0; font-size: 16px; color: #f3f4f6; } | |
| #footer-info ul { padding-left: 20px; color: #d1d5db; font-size: 13.5px; line-height: 1.6; } | |
| """ | |
| EJEMPLOS = [ | |
| ["BREAKING: 7.2 magnitude earthquake hits coastal region, buildings collapsed in downtown area"], | |
| ["Our thoughts and prayers go out to everyone affected by the hurricane"], | |
| ["Alerta sanitaria oficial en Iquique tras sismo: no consumir agua de la red"], | |
| ["Que tristeza lo que pasa en Venezuela. #FuerzaVenezuela"], | |
| ["Rescue teams working to free 20 people trapped in collapsed hotel in downtown Port-au-Prince"], | |
| ["Peanut butter jelly sandwiches are the best"], | |
| ["RT @24HorasTVN: Imagen del conductor del tren accidentado en Espana da la vuelta al mundo"], | |
| ["Happy birthday bro @Hurricane_Dolce"], | |
| ] | |
| with gr.Blocks(theme=gr.themes.Soft(primary_hue="red", secondary_hue="orange", neutral_hue="slate"), | |
| css=CUSTOM_CSS, | |
| title="CrisisBench Classifier") as app: | |
| gr.HTML(""" | |
| <div id="header-banner"> | |
| <h1>🚨 Filtro de mensajes en emergencias</h1> | |
| <p>Clasifica tuits de desastres como <strong>informativos</strong> (datos operativos) o <strong>no informativos</strong> (opiniones, ruido). Compara dos modelos en vivo.</p> | |
| </div> | |
| """) | |
| with gr.Row(): | |
| with gr.Column(scale=5): | |
| entrada = gr.Textbox( | |
| label="📝 Mensaje a clasificar", | |
| placeholder="Pega aqui un tuit, un mensaje de Facebook o un WhatsApp sobre un desastre...", | |
| lines=5, | |
| ) | |
| boton = gr.Button("Clasificar mensaje", variant="primary", size="lg") | |
| gr.Examples( | |
| examples=EJEMPLOS, | |
| inputs=entrada, | |
| label="Prueba con estos ejemplos:", | |
| examples_per_page=8, | |
| ) | |
| with gr.Column(scale=6): | |
| texto_limpio = gr.Textbox( | |
| label="🧼 Texto despues del preprocesamiento", | |
| interactive=False, | |
| lines=2, | |
| ) | |
| with gr.Row(): | |
| with gr.Column(): | |
| gr.HTML('<div class="model-title">XLM-RoBERTa (fine-tuned)</div><div class="model-subtitle">Transformer multilingue preentrenado · 94 idiomas · F1 macro test: 0.873</div>') | |
| out_xlm = gr.HTML("<div style='color:#9ca3af; padding:16px;'>Esperando mensaje...</div>") | |
| with gr.Column(): | |
| gr.HTML('<div class="model-title">Regresion Logistica + TF-IDF</div><div class="model-subtitle">Modelo clasico rapido · F1 macro test: 0.856 · Costo cero</div>') | |
| out_lr = gr.HTML("<div style='color:#9ca3af; padding:16px;'>Esperando mensaje...</div>") | |
| acuerdo = gr.HTML("") | |
| gr.HTML(""" | |
| <div id="footer-info"> | |
| <h3>Sobre esta aplicacion</h3> | |
| <p style="color:#d1d5db; font-size:13.5px; line-height:1.6;"> | |
| Proyecto final de <strong>Tecnicas de Aprendizaje de Maquina</strong> · Pontificia Universidad Javeriana 2026-2 · Dataset <a href="https://huggingface.co/datasets/timhigins/crisisbench" style="color:#fca5a5;">timhigins/crisisbench</a> (166k mensajes, 61 eventos de desastre, 7 idiomas). | |
| </p> | |
| <h3>Limitaciones conocidas</h3> | |
| <ul> | |
| <li><strong>Calidad de etiquetas</strong>: una revision manual de 100 mensajes mostro 18% de desacuerdo con la etiqueta oficial. Los modelos aprenden ese ruido. Por eso la app muestra un aviso cuando detecta un caso tipico de etiqueta dudosa.</li> | |
| <li><strong>Espanol</strong>: solo 3% del corpus, con 32% de etiquetas discutibles en la clase "informativo". Precision menor que en ingles.</li> | |
| <li><strong>Eventos nuevos</strong>: XLM-R generaliza mejor (caida de 9 puntos F1 en evento Chile dejado fuera) que LogReg (caida de 22 puntos).</li> | |
| <li><strong>No reemplaza a un humano</strong>: es un filtro de primer paso para reducir el volumen a revisar en una emergencia.</li> | |
| </ul> | |
| </div> | |
| """) | |
| boton.click(clasificar, inputs=entrada, | |
| outputs=[texto_limpio, out_xlm, out_lr, acuerdo]) | |
| if __name__ == "__main__": | |
| app.launch(server_name="0.0.0.0", server_port=7860, ssr_mode=False) | |