synthetic data generator

week3/community-contributions/week3_assignment_data_generator_congress.py

@@ -0,0 +1,307 @@
+# %% [markdown]
+''' Este script permite **generar datos sintéticos realistas** para analizar la cobertura mediática de eventos (congresos, conferencias, foros) utilizando **modelos de lenguaje avanzados** (GPT-4 y Claude).
+### Tres tipos de datasets generables:
+
+1. **Eventos**
+   - Información básica del evento: nombre, fechas, ubicación, asistencia esperada
+   - Ejemplo: "Congreso de Turismo Sostenible 2025 en Valencia"
+
+2. **Cobertura Mediática**
+   - Artículos de prensa sobre el evento
+   - Incluye: título, resumen, fuente, sentimiento, alcance
+   - Fuentes variadas: nacionales, regionales, internacionales, especializadas
+
+3. **Menciones en Redes Sociales**
+   - Posts de Twitter, Facebook, Instagram, LinkedIn, TikTok
+   - Métricas de engagement: likes, shares, comentarios, alcance
+   - Análisis de sentimiento y hashtags
+'''
+
+
+import os
+import json
+import tempfile
+from datetime import datetime, timedelta
+from dateutil import parser as dateparser
+from dotenv import load_dotenv
+
+import pandas as pd
+import gradio as gr
+from faker import Faker
+import openai
+from anthropic import Anthropic
+
+# Configuración inicial
+load_dotenv()
+fake = Faker()
+
+OPENAI_API_KEY = os.getenv("OPENAI_API_KEY")
+ANTHROPIC_API_KEY = os.getenv("ANTHROPIC_API_KEY")
+MAX_RECORDS_ALLOWED = 5
+
+# Configuración de clientes
+openai.api_key = OPENAI_API_KEY
+anthropic_client = Anthropic(api_key=ANTHROPIC_API_KEY)
+
+# ============================================================================
+# UTILIDADES
+# ============================================================================
+
+def extract_json_from_text(text: str) -> str:
+    """Extrae JSON array de respuesta LLM, manejando bloques ```json```."""
+    if not text:
+        raise ValueError("Respuesta vacía del modelo")
+
+    # Remover bloques de código
+    if "```json" in text:
+        text = text.split("```json", 1)[1].split("```", 1)[0]
+    elif "```" in text:
+        text = text.split("```", 1)[1].split("```", 1)[0]
+
+    # Buscar inicio del array
+    idx = text.find('[')
+    if idx >= 0:
+        text = text[idx:]
+
+    return text.strip()
+
+def parse_json_safe(text: str) -> list:
+    """Parsea JSON array con manejo robusto de errores."""
+    try:
+        json_str = extract_json_from_text(text)
+        data = json.loads(json_str)
+        if not isinstance(data, list):
+            raise ValueError("El JSON no es un array")
+        return data
+    except Exception as e:
+        snippet = text[:500] if text else ""
+        raise ValueError(f"Error parseando JSON: {e}\nSnippet: {snippet}")
+
+def save_to_csv(df: pd.DataFrame) -> str:
+    """Guarda DataFrame en archivo temporal CSV."""
+    fd, path = tempfile.mkstemp(suffix=".csv", prefix="synthetic_")
+    os.close(fd)
+    df.to_csv(path, index=False)
+    return path
+
+# ============================================================================
+# PROMPTS
+# ============================================================================
+
+PROMPTS = {
+    "events": """Genera exactamente {n} eventos en formato JSON array. Cada objeto debe tener:
+- event_id: string (EVT-XXXXXX)
+- event_name: string
+- start_date: ISO 8601
+- end_date: ISO 8601
+- location_city: string
+- main_topic: string
+- expected_attendance: integer
+
+Contexto: Tipo={event_type}, Fechas={start_iso} a {end_iso}, Location={location}, Asistencia={attendance}, Idioma={language}
+
+Devuelve SOLO el array JSON, sin texto adicional.""",
+
+    "media_coverage": """Genera exactamente {n} artículos en formato JSON array. Cada objeto debe tener:
+- article_id, event_id, published_at (ISO 8601), source_name, source_type (national/regional/international)
+- title, summary (2 oraciones), language, sentiment_score (-1 a 1), stance (positive/neutral/negative)
+- reach (integer), topic_tags (array), entities (array)
+
+Contexto: Evento={event_name} ({event_id}), Tipo={event_type}, Fechas={start_iso} a {end_iso}, Location={location}, Source={source_type}, Idioma={language}
+
+Devuelve SOLO el array JSON con variedad y coherencia temporal.""",
+
+    "social_mentions": """Genera exactamente {n} posts en formato JSON array. Cada objeto debe tener:
+- post_id, event_id, timestamp (ISO 8601), platform (twitter/facebook/instagram/linkedin/tiktok), handle
+- text, language, sentiment_score (-1 a 1), stance (positive/neutral/negative)
+- likes, shares, replies, reach (integers), hashtags (array), entities (array)
+
+Contexto: Evento={event_name} ({event_id}), Tipo={event_type}, Fechas={start_iso} a {end_iso}, Location={location}, Idioma={language}
+
+Devuelve SOLO el array JSON con variedad en plataformas y engagement natural."""
+}
+
+# ============================================================================
+# LLAMADAS A MODELOS
+# ============================================================================
+
+def call_openai(prompt: str, api_key: str, model: str = "gpt-4o-mini", temp: float = 0.7) -> tuple:
+    """Llama a OpenAI y extrae texto de respuesta."""
+    try:
+        client = openai.OpenAI(api_key=api_key)
+        response = client.chat.completions.create(
+            model=model,
+            messages=[{"role": "user", "content": prompt}],
+            max_tokens=2000,
+            temperature=temp
+        )
+        return response.choices[0].message.content, None
+    except Exception as e:
+        return None, f"OpenAI error: {e}"
+
+def call_anthropic(prompt: str, api_key: str, model: str = "claude-3-5-haiku-latest", temp: float = 0.7) -> tuple:
+    """Llama a Anthropic y extrae texto de respuesta."""
+    try:
+        client = Anthropic(api_key=api_key)
+        response = client.messages.create(
+            model=model,
+            max_tokens=2000,
+            temperature=temp,
+            messages=[{"role": "user", "content": prompt}]
+        )
+        return "".join(block.text for block in response.content if hasattr(block, "text")), None
+    except Exception as e:
+        return None, f"Anthropic error: {e}"
+
+# ============================================================================
+# GENERACIÓN
+# ============================================================================
+
+def generate_dataset(provider: str, dataset_type: str, n: int, context: dict, 
+                    language: str, api_key: str = None, temp: float = 0.7) -> tuple:
+    """Genera dataset sintético usando LLM especificado."""
+
+    if not (0 < n <= MAX_RECORDS_ALLOWED):
+        raise ValueError(f"n debe estar entre 1 y {MAX_RECORDS_ALLOWED}")
+
+    # Preparar prompt
+    prompt = PROMPTS[dataset_type].format(n=n, language=language, **context)
+
+    # Seleccionar modelo y API key
+    models = {"openai": "gpt-4o-mini", "anthropic": "claude-3-5-haiku-latest"}
+    model = models.get(provider)
+
+    if not api_key:
+        api_key = OPENAI_API_KEY if provider == "openai" else ANTHROPIC_API_KEY
+
+    # Llamar modelo
+    call_fn = call_openai if provider == "openai" else call_anthropic
+    text, err = call_fn(prompt, api_key, model, temp)
+
+    if err:
+        return None, err
+
+    # Parsear respuesta
+    try:
+        records = parse_json_safe(text)
+        return records, None
+    except Exception as e:
+        return None, str(e)
+
+# ============================================================================
+# INTERFAZ GRADIO
+# ============================================================================
+
+def run_generation(provider, event_type, start_date, end_date, location, 
+                  attendance, source_type, language, dataset_type, n_items, api_key_input):
+    """Handler principal para generación desde UI."""
+
+    try:
+        n_items = int(n_items)
+        if not (0 < n_items <= MAX_RECORDS_ALLOWED):
+            return "", f"Número debe estar entre 1 y {MAX_RECORDS_ALLOWED}", None
+
+        # Parsear fechas
+        start_dt = dateparser.parse(start_date)
+        end_dt = dateparser.parse(end_date)
+
+        # Preparar contexto
+        event_id = f"EVT-{fake.unique.random_number(digits=6)}"
+        context = {
+            "event_id": event_id,
+            "event_name": f"{event_type} - {location}",
+            "start_iso": start_dt.isoformat(),
+            "end_iso": end_dt.isoformat(),
+            "location": location,
+            "attendance": int(attendance),
+            "event_type": event_type,
+            "source_type": source_type
+        }
+
+        # Mapear idioma
+        lang_map = {"español": "español", "ingles": "english"}
+        lang = lang_map.get(language.lower(), language)
+
+        # Generar datos
+        records, err = generate_dataset(
+            provider=provider.lower(),
+            dataset_type=dataset_type,
+            n=n_items,
+            context=context,
+            language=lang,
+            api_key=api_key_input.strip() or None
+        )
+
+        if err:
+            return "", f"Error: {err}", None
+
+        # Crear DataFrame y guardar
+        df = pd.DataFrame(records)
+        csv_path = save_to_csv(df)
+        preview = df.head(20).to_html(index=False)
+        status = f"✅ Generados {len(df)} registros con {provider} ({lang})"
+
+        return preview, status, csv_path
+
+    except Exception as e:
+        return "", f"Error: {str(e)}", None
+
+# ============================================================================
+# UI
+# ============================================================================
+
+with gr.Blocks(title="Generador Sintético de Datos") as demo:
+    gr.Markdown("# Generador Sintético de Datos Mediáticos de Eventos")
+
+    with gr.Row():
+        with gr.Column(scale=2):
+            gr.Markdown("### Configuración")
+            provider = gr.Radio(["openai", "anthropic"], value="openai", label="Proveedor LLM")
+            api_key_input = gr.Textbox(label="API Key (opcional)", type="password", 
+                                      placeholder="Deja vacío para usar .env")
+
+            gr.Markdown("### Parámetros del Evento")
+            event_type = gr.Dropdown(
+                ["Congreso de odontólogos", "Congreso de diabetes", "Congreso de energía", 
+                 "Congreso de movilidad urbana", "Foro empresarial"],
+                value="Congreso de turismo", label="Tipo de evento"
+            )
+            start_date = gr.Textbox(value=datetime.now().date().isoformat(), label="Fecha inicio (YYYY-MM-DD)")
+            end_date = gr.Textbox(value=(datetime.now() + timedelta(days=2)).date().isoformat(), 
+                                 label="Fecha fin (YYYY-MM-DD)")
+            location = gr.Dropdown(
+                ["Valencia", "Sydney", "Ljubljana", "Larissa", "Heidelberg", "Paris", "London", "New York"],
+                value="Valencia", label="Ciudad"
+            )
+            attendance = gr.Number(value=300, label="Asistencia esperada", precision=0)
+
+            gr.Markdown("### Parámetros de Generación")
+            source_type = gr.Radio(
+                ["national", "regional", "international"],
+                value="national", label="Tipo de fuente (media)"
+            )
+            language = gr.Radio(["español", "ingles"], value="español", label="Idioma")
+            dataset_type = gr.Radio(
+                ["events", "media_coverage", "social_mentions"],
+                value="media_coverage", label="Tipo de dataset"
+            )
+            n_items = gr.Slider(1, MAX_RECORDS_ALLOWED, value=5, step=1, 
+                              label=f"Cantidad de registros (máx {MAX_RECORDS_ALLOWED})")
+
+            generate_btn = gr.Button("Generar Dataset", variant="primary")
+
+        with gr.Column(scale=3):
+            gr.Markdown("### Resultados")
+            preview = gr.HTML("<i>Preview aparecerá aquí tras generar...</i>")
+            status = gr.Textbox(label="Estado", interactive=False)
+            csv_file = gr.File(label="Descargar CSV")
+
+    generate_btn.click(
+        fn=run_generation,
+        inputs=[provider, event_type, start_date, end_date, location, attendance, 
+                source_type, language, dataset_type, n_items, api_key_input],
+        outputs=[preview, status, csv_file]
+    )
+
+if __name__ == "__main__":
+    demo.launch(share=False, debug=True)