Upload IVRIUS_notebook2.ipynb with huggingface_hub

IVRIUS_notebook2.ipynb

@@ -32,10 +32,7 @@
    "cell_type": "code",
    "metadata": {},
    "source": [
-    "# ── Célula 1: Instalar dependências ──────────────────────────────\n",
-    "!pip install -q transformers accelerate bitsandbytes sentencepiece\n",
-    "!pip install -q sentence-transformers datasets huggingface_hub\n",
-    "print('✅ Dependências instaladas!')"
+    "# ── Célula 1: Instalar dependências ──────────────────────────────\n!pip install -q llama-cpp-python --extra-index-url https://abetlen.github.io/llama-cpp-python/whl/cu121\n!pip install -q sentence-transformers datasets huggingface_hub\nprint('✅ Dependências instaladas!')"
    ],
    "execution_count": null,
    "outputs": []
@@ -62,7 +59,7 @@
    "cell_type": "code",
    "metadata": {},
    "source": [
-    "# ── Célula 3: Carregar Jurema-7B (modelo jurídico BR) ────────────\nfrom transformers import AutoTokenizer, AutoModelForCausalLM, pipeline\nimport torch\n\nprint('📥 Carregando Jurema-7B...')\nprint('   (pode demorar 5-10 min na primeira vez)')\n\n# Jurema sem gate — GGUF via llama.cpp\nMODEL_ID = 'mauroneto/Jurema-7B-Q4_K_M-GGUF'\n# Alternativas:\n# MODEL_ID = 'rlmoura/Jurema-7B-Q4_K_M-GGUF'\n# MODEL_ID = 'GenivalFeitoza73/qwen32b-juridico'\n\ntokenizer = AutoTokenizer.from_pretrained(MODEL_ID, token=HF_TOKEN)\nmodel_jurema = AutoModelForCausalLM.from_pretrained(\n    MODEL_ID,\n    torch_dtype=torch.float16,\n    device_map='auto',\n    load_in_4bit=True,\n    gguf_file='Jurema-7B-Q4_K_M.gguf',\n    token=HF_TOKEN\n)\n\npipe = pipeline(\n    'text-generation',\n    model=model_jurema,\n    tokenizer=tokenizer,\n    max_new_tokens=512,\n    temperature=0.3,\n    do_sample=True,\n    pad_token_id=tokenizer.eos_token_id\n)\nprint('✅ Jurema-7B carregado!')"
+    "# ── Célula 3: Carregar Jurema-7B GGUF via llama-cpp ─────────────\nfrom llama_cpp import Llama\nfrom huggingface_hub import hf_hub_download\n\nprint('📥 Baixando Jurema-7B GGUF...')\n\n# Baixa o GGUF do HuggingFace\ngguf_path = hf_hub_download(\n    repo_id=\"MMR408/proejto2\",\n    filename=\"jurema-7b-q4_k_m.gguf\",\n    repo_type=\"dataset\",\n    token=HF_TOKEN\n)\nprint(f'✅ Download completo: {gguf_path}')\n\nprint('📥 Carregando Jurema-7B na memória...')\nmodel_jurema = Llama(\n    model_path=gguf_path,\n    n_gpu_layers=-1,   # usa GPU toda\n    n_ctx=2048,\n    verbose=False\n)\n\ndef pipe(prompt_text):\n    output = model_jurema(\n        prompt_text,\n        max_tokens=512,\n        temperature=0.3,\n        stop=[\"```\", \"\\n\\n\\n\"]\n    )\n    return [{\"generated_text\": prompt_text + output[\"choices\"][0][\"text\"]}]\n\nprint('✅ Jurema-7B carregado!')"
    ],
    "execution_count": null,
    "outputs": []