Pipeline: E5-Large-Instruct → BERTopic (UMAP 10d, HDBSCAN mcs=30/ms=5, KeyBERT+MMR) → Marco-Mini-Instruct i1-Q6_K
\n", "{summary_html}\n", "{tabs_html}\n", "{panels_html}\n", "{js}\n", "\n", "\"\"\"\n", "\n", "dashboard_path = Path(\"/content/drive/MyDrive/Compartilhados/community_notes_pt/sessao_bertopic/dashboard_community_notes.html\")\n", "dashboard_path.write_text(full_html, encoding=\"utf-8\")\n", "print(f\"Dashboard salvo: {dashboard_path}\")\n", "print(f\"Tamanho: {dashboard_path.stat().st_size / 1e6:.1f} MB\")\n", "print(f\"Abas: {', '.join(name for name, _ in figs)}\")" ], "metadata": { "colab": {}, "id": "ckjgfzR-3nRh" }, "id": "ckjgfzR-3nRh", "execution_count": null, "outputs": [] }, { "cell_type": "code", "source": [ "from google.colab import drive\n", "from pathlib import Path\n", "from bertopic import BERTopic\n", "import numpy as np\n", "import pandas as pd\n", "\n", "drive.mount(\"/content/drive\")\n", "\n", "BASE = Path(\"/content/drive/MyDrive/Compartilhados/community_notes_pt/sessao_bertopic\")\n", "SESSION = BASE / \"marco_mini_i1\"\n", "E5_SESSION = SESSION / \"e5_instruct\"\n", "\n", "# ── 1. Embeddings e5-instruct ──\n", "embeddings = np.load(SESSION / \"embeddings_tagged_e5-instruct.npy\")\n", "print(f\"Embeddings: {embeddings.shape}\")\n", "\n", "# ── 2. Modelo BERTopic (e5-instruct, 898 tópicos) ──\n", "topic_model = BERTopic.load(str(E5_SESSION / \"model\"))\n", "topics = topic_model.topics_\n", "print(f\"Modelo: {len(set(topics))-1} tópicos\")\n", "\n", "# ── 3. Metadados dos tópicos ──\n", "topic_meta_df = pd.read_parquet(E5_SESSION / \"topic_metadata_marco_mini.parquet\")\n", "topic_metadata = {\n", " int(row[\"topic\"]): {\n", " \"contexto_central\": row.get(\"contexto_central\", \"\"),\n", " \"categoria_ampla\": row.get(\"categoria_ampla\", \"Outro\"),\n", " \"rotulo_curto\": row.get(\"rotulo_curto\", \"\"),\n", " }\n", " for _, row in topic_meta_df.iterrows()\n", "}\n", "print(f\"Metadados: {len(topic_metadata)} tópicos\")\n", "\n", "# ── 4. Documentos anotados ──\n", "df = pd.read_parquet(E5_SESSION / \"notas_com_topicos_v3.parquet\")\n", "docs = df[\"doc\"].tolist()\n", "df_topics = df[df[\"topic\"] != -1].copy()\n", "print(f\"Docs: {len(df):,} | Com tópico: {len(df_topics):,}\")\n", "\n", "# ── 5. Macrotemas ──\n", "macro_summary = pd.read_csv(SESSION / \"macro_summary.csv\")\n", "macro_labels = {\n", " int(row[\"macrotheme_id\"]): {\n", " \"rotulo_macro\": row[\"rotulo_macro\"],\n", " \"categoria_dominante\": row[\"categoria_dominante\"],\n", " }\n", " for _, row in macro_summary.iterrows()\n", "}\n", "macro_label_map = {mid: v[\"rotulo_macro\"] for mid, v in macro_labels.items()}\n", "macro_cat_map = {mid: v[\"categoria_dominante\"] for mid, v in macro_labels.items()}\n", "print(f\"Macrotemas: {len(macro_labels)}\")\n", "\n", "# ── 6. Reconstruir topic_info_full ──\n", "valid_topic_ids = sorted(t for t in set(topics) if t != -1)\n", "topic_info_full = topic_model.get_topic_info().copy()\n", "topic_info_full = topic_info_full[topic_info_full[\"Topic\"] != -1].copy()\n", "\n", "tid_to_macro = df_topics.drop_duplicates(\"topic\").set_index(\"topic\")[\"macrotheme_id\"].to_dict()\n", "topic_info_full[\"macrotheme_id\"] = topic_info_full[\"Topic\"].map(tid_to_macro)\n", "topic_info_full[\"rotulo_curto\"] = topic_info_full[\"Topic\"].map(\n", " {t: topic_metadata.get(t, {}).get(\"rotulo_curto\", \"\") for t in valid_topic_ids}\n", ")\n", "\n", "# ── 7. Embeddings dos tópicos ──\n", "if hasattr(topic_model, \"topic_embeddings_\") and topic_model.topic_embeddings_ is not None:\n", " raw_emb = np.asarray(topic_model.topic_embeddings_)\n", " topic_info_tmp = topic_model.get_topic_info()\n", " tid_to_idx = dict(zip(topic_info_tmp[\"Topic\"], range(len(topic_info_tmp))))\n", " topic_embs = np.vstack([raw_emb[tid_to_idx[t]] for t in valid_topic_ids])\n", "else:\n", " centroids = []\n", " for tid in valid_topic_ids:\n", " mask = np.array(topics) == tid\n", " c = embeddings[mask].mean(axis=0)\n", " c /= max(np.linalg.norm(c), 1e-9)\n", " centroids.append(c)\n", " topic_embs = np.vstack(centroids)\n", "print(f\"Topic embeddings: {topic_embs.shape}\")\n", "\n", "# ── 8. Procedurais ──\n", "procedural_keywords = [\"comunidade\", \"comentário\", \"opinião pessoal\", \"nota da comunidade\",\n", " \"communitynotes\", \"sátira sem\", \"resposta em comentário\",\n", " \"contribuição da comunidade\", \"guia de notas\", \"uso das notas\"]\n", "\n", "for _, row in topic_meta_df.iterrows():\n", " tid = int(row[\"topic\"])\n", " text = f\"{row.get('rotulo_curto','')} {row.get('contexto_central','')}\".lower()\n", " if any(kw in text for kw in procedural_keywords):\n", " if tid in topic_metadata:\n", " topic_metadata[tid][\"categoria_ampla\"] = \"Procedural\"\n", "\n", "proc_macros = [mid for mid, cat in macro_cat_map.items() if cat == \"Procedural\"]\n", "proc_topics_kw = {tid for tid, m in topic_metadata.items() if m[\"categoria_ampla\"] == \"Procedural\"}\n", "proc_topics_macro = set(topic_info_full[topic_info_full[\"macrotheme_id\"].isin(proc_macros)][\"Topic\"])\n", "proc_topics_combined = proc_topics_kw | proc_topics_macro\n", "\n", "df_substantivo = df_topics[~df_topics[\"topic\"].isin(proc_topics_combined)].copy()\n", "topic_info_substantivo = topic_info_full[~topic_info_full[\"Topic\"].isin(proc_topics_combined)].copy()\n", "\n", "print(f\"Procedurais: {len(proc_topics_combined)} tópicos | {len(proc_macros)} macros\")\n", "print(f\"Substantivos: {len(df_substantivo):,} docs\")\n", "\n", "# ── 9. Config ──\n", "GRAPH_K = 6\n", "GRAPH_MIN_SEM_SIM = 0.60\n", "TOP_N_WORDS = 10\n", "\n", "print(\"\\nTudo carregado. Rode o bloco do dashboard.\")" ], "metadata": { "colab": {}, "id": "sHCJ1Mdk7Zy6" }, "id": "sHCJ1Mdk7Zy6", "execution_count": null, "outputs": [] } ] }