{
 "cells": [
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "# Quickstart — Community Notes Reranker (PT-BR)\n",
    "\n",
    "Notebook mínimo de inferência. Baixa o modelo (base Qwen3-Reranker-0.6B + adapter LoRA), monta o template e devolve a probabilidade de utilidade para um par `(tweet, nota)`.\n",
    "\n",
    "- **Runtime sugerido:** GPU (T4 basta). CPU também funciona, mas ~5-10s por inferência.\n",
    "- **Modos disponíveis:** *fold único* (rápido) e *ensemble dos 5 folds* (reproduz exatamente o número reportado no model card).\n"
   ]
  },
  {
   "cell_type": "code",
   "metadata": {},
   "source": [
    "# Instala dependencias se necessario\n",
    "import sys, subprocess, importlib\n",
    "for mod, pkg in [(\"torch\",\"torch\"), (\"transformers\",\"transformers\"),\n",
    "                 (\"peft\",\"peft\"), (\"huggingface_hub\",\"huggingface_hub\")]:\n",
    "    try: importlib.import_module(mod)\n",
    "    except Exception:\n",
    "        subprocess.run([sys.executable, \"-m\", \"pip\", \"install\", \"-q\", pkg], check=True)\n"
   ],
   "outputs": [],
   "execution_count": null
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## Carrega base + um fold (modo rápido)"
   ]
  },
  {
   "cell_type": "code",
   "metadata": {},
   "source": [
    "import json, torch\n",
    "from transformers import AutoTokenizer, AutoModelForCausalLM\n",
    "from peft import PeftModel\n",
    "from huggingface_hub import snapshot_download\n",
    "\n",
    "REPO = \"histlearn/community-notes-reranker-ptbr\"\n",
    "path = snapshot_download(REPO, allow_patterns=[\"manifesto.json\", \"adapter_fold_1/*\"])\n",
    "m    = json.load(open(f\"{path}/manifesto.json\"))\n",
    "\n",
    "tok  = AutoTokenizer.from_pretrained(m[\"base_model\"], padding_side=\"left\")\n",
    "dtype = torch.float16 if torch.cuda.is_available() else torch.float32\n",
    "base  = AutoModelForCausalLM.from_pretrained(m[\"base_model\"], torch_dtype=dtype)\n",
    "model = PeftModel.from_pretrained(base, f\"{path}/adapter_fold_1\")\n",
    "if torch.cuda.is_available(): model.cuda()\n",
    "model.eval()\n",
    "print(f\"Modelo pronto em: {model.device}\")\n"
   ],
   "outputs": [],
   "execution_count": null
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## Função de inferência"
   ]
  },
  {
   "cell_type": "code",
   "metadata": {},
   "source": [
    "def util_prob(tweet: str, nota: str) -> float:\n",
    "    \"\"\"Probabilidade de que a comunidade marcaria a nota como util.\n",
    "    Threshold otimo medido sob CV (Platt scaling) = 0.38.\"\"\"\n",
    "    text = (m[\"prompt_prefixo\"] + \"<Instruct>: \" + m[\"instrucao\"] +\n",
    "            \"\\n<Query>: \" + tweet + \"\\n<Document>: \" + nota + m[\"prompt_sufixo\"])\n",
    "    enc = tok(text, return_tensors=\"pt\", truncation=True, max_length=m[\"max_length\"]).to(model.device)\n",
    "    with torch.no_grad():\n",
    "        logits = model(**enc).logits[:, -1, :]\n",
    "    return float(torch.sigmoid(logits[:, m[\"id_yes\"]] - logits[:, m[\"id_no\"]]).item())\n"
   ],
   "outputs": [],
   "execution_count": null
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## Exemplo"
   ]
  },
  {
   "cell_type": "code",
   "metadata": {},
   "source": [
    "tweet = (\"Lula anunciou que o salario minimo subira para R$ 5 mil em 2026.\")\n",
    "nota  = (\"E falso. Em 12/12/2024, o presidente Lula anunciou que o salario minimo \"\n",
    "         \"subiria de R$ 1.412 para R$ 1.518 a partir de janeiro de 2025, segundo a \"\n",
    "         \"Agencia Brasil (https://agenciabrasil.ebc.com.br). Nao ha qualquer \"\n",
    "         \"anuncio oficial de valor proximo a R$ 5 mil.\")\n",
    "\n",
    "p = util_prob(tweet, nota)\n",
    "print(f\"P(util) = {p:.4f}\")\n",
    "print(f\"Classificacao (threshold 0.38): {'UTIL' if p >= 0.38 else 'NAO-UTIL'}\")\n"
   ],
   "outputs": [],
   "execution_count": null
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## Ensemble dos 5 folds (reproduz o número do model card)\n",
    "\n",
    "Para resultados estatisticamente comparáveis aos reportados (macro-F1 0.7920), use a média das probabilidades dos 5 adapters."
   ]
  },
  {
   "cell_type": "code",
   "metadata": {},
   "source": [
    "from huggingface_hub import snapshot_download\n",
    "path_full = snapshot_download(REPO, allow_patterns=[\"manifesto.json\", \"adapter_fold_*/*\"])\n",
    "\n",
    "def util_prob_ensemble(tweet: str, nota: str) -> float:\n",
    "    probs = []\n",
    "    for k in range(1, 6):\n",
    "        m_k = PeftModel.from_pretrained(base, f\"{path_full}/adapter_fold_{k}\")\n",
    "        m_k.eval()\n",
    "        text = (m[\"prompt_prefixo\"] + \"<Instruct>: \" + m[\"instrucao\"] +\n",
    "                \"\\n<Query>: \" + tweet + \"\\n<Document>: \" + nota + m[\"prompt_sufixo\"])\n",
    "        enc = tok(text, return_tensors=\"pt\", truncation=True, max_length=m[\"max_length\"]).to(m_k.device)\n",
    "        with torch.no_grad():\n",
    "            l = m_k(**enc).logits[:, -1, :]\n",
    "        probs.append(float(torch.sigmoid(l[:, m[\"id_yes\"]] - l[:, m[\"id_no\"]]).item()))\n",
    "        # Libera o adapter k antes de carregar o k+1\n",
    "        if hasattr(m_k, \"unload\"):\n",
    "            m_k.unload()\n",
    "    return sum(probs) / 5\n",
    "\n",
    "p_ens = util_prob_ensemble(tweet, nota)\n",
    "print(f\"P(util) ensemble = {p_ens:.4f}\")\n"
   ],
   "outputs": [],
   "execution_count": null
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## Próximos passos\n",
    "\n",
    "- Para documentação completa, métricas e contexto do projeto, ver o [model card](https://huggingface.co/histlearn/community-notes-reranker-ptbr).\n",
    "- Para reproduzir o treino fold-a-fold ou regenerar os artefatos, ver o notebook `02_pipeline_experimento.ipynb` no [Space do projeto](https://huggingface.co/spaces/histlearn/communitynotesbr).\n",
    "- Para o dataset bruto, ver [`histlearn/notas-comunidade-ptbr`](https://huggingface.co/datasets/histlearn/notas-comunidade-ptbr).\n"
   ]
  }
 ],
 "metadata": {
  "kernelspec": {
   "display_name": "Python 3",
   "language": "python",
   "name": "python3"
  },
  "language_info": {
   "name": "python",
   "version": "3.11"
  }
 },
 "nbformat": 4,
 "nbformat_minor": 5
}