copaca-training.ipynb

{
  "nbformat": 4,
  "nbformat_minor": 0,
  "metadata": {
    "colab": {
      "provenance": [],
      "gpuType": "T4",
      "machine_shape": "hm"
    },
    "kernelspec": {
      "name": "python3",
      "display_name": "Python 3"
    },
    "language_info": {
      "name": "python"
    },
    "accelerator": "GPU"
  },
  "cells": [
    {
      "cell_type": "markdown",
      "source": [
        "# 패키지 다운로드"
      ],
      "metadata": {
        "id": "I1aM0iADTje0"
      }
    },
    {
      "cell_type": "code",
      "execution_count": null,
      "metadata": {
        "id": "FuXIFTFapAMI"
      },
      "outputs": [],
      "source": [
        "!pip install -q -U bitsandbytes\n",
        "!pip install -q -U git+https://github.com/huggingface/transformers.git \n",
        "!pip install -q -U git+https://github.com/huggingface/peft.git\n",
        "!pip install -q -U git+https://github.com/huggingface/accelerate.git\n",
        "!pip install -q datasets"
      ]
    },
    {
      "cell_type": "markdown",
      "source": [
        "# 데이터 셋 가져오기"
      ],
      "metadata": {
        "id": "_9fHSqQHTqkU"
      }
    },
    {
      "cell_type": "code",
      "source": [
        "from datasets import load_dataset\n",
        "\n",
        "data = load_dataset(\"paragonnov/coway_faq\")"
      ],
      "metadata": {
        "id": "jm4FzCvfeYcK"
      },
      "execution_count": null,
      "outputs": []
    },
    {
      "cell_type": "code",
      "source": [
        "data"
      ],
      "metadata": {
        "id": "2KUhV7x3e6Db"
      },
      "execution_count": null,
      "outputs": []
    },
    {
      "cell_type": "markdown",
      "source": [
        "# 학습을 위한 *text* 데이터 매핑"
      ],
      "metadata": {
        "id": "W-WnBJsbTvG5"
      }
    },
    {
      "cell_type": "code",
      "source": [
        "data = data.map(\n",
        "    lambda x: {'text': f\"### 질문: {x['instruction']}\\n\\n### 답변: {x['output']}<|endoftext|>\" }\n",
        ")"
      ],
      "metadata": {
        "id": "0FbgsI9sezTJ"
      },
      "execution_count": null,
      "outputs": []
    },
    {
      "cell_type": "code",
      "source": [
        "data"
      ],
      "metadata": {
        "id": "w7-HuF1p6EQT"
      },
      "execution_count": null,
      "outputs": []
    },
    {
      "cell_type": "markdown",
      "source": [
        "# 모델 조각 불러오기"
      ],
      "metadata": {
        "id": "dreUUUrdT2-1"
      }
    },
    {
      "cell_type": "code",
      "source": [
        "import torch\n",
        "from transformers import AutoTokenizer, AutoModelForCausalLM, BitsAndBytesConfig\n",
        "\n",
        "model_id = \"beomi/polyglot-ko-12.8b-safetensors\"  # safetensors 컨버팅된 레포\n",
        "bnb_config = BitsAndBytesConfig(\n",
        "    load_in_4bit=True,\n",
        "    bnb_4bit_use_double_quant=True,\n",
        "    bnb_4bit_quant_type=\"nf4\",\n",
        "    bnb_4bit_compute_dtype=torch.bfloat16\n",
        ")\n",
        "\n",
        "tokenizer = AutoTokenizer.from_pretrained(model_id)\n",
        "model = AutoModelForCausalLM.from_pretrained(model_id, quantization_config=bnb_config, device_map={\"\":0})"
      ],
      "metadata": {
        "id": "E0Nl5mWL0k2T"
      },
      "execution_count": null,
      "outputs": []
    },
    {
      "cell_type": "code",
      "source": [
        "data = data.map(lambda samples: tokenizer(samples[\"text\"]), batched=True)"
      ],
      "metadata": {
        "id": "C4TDUgDbhyhK"
      },
      "execution_count": null,
      "outputs": []
    },
    {
      "cell_type": "code",
      "source": [
        "data['train'][0]['text']"
      ],
      "metadata": {
        "id": "h61VdWpSAJEp"
      },
      "execution_count": null,
      "outputs": []
    },
    {
      "cell_type": "markdown",
      "source": [
        "# PEFT 로 LB 학습"
      ],
      "metadata": {
        "id": "EXROYJDNT9DJ"
      }
    },
    {
      "cell_type": "code",
      "source": [
        "from peft import prepare_model_for_kbit_training\n",
        "\n",
        "model.gradient_checkpointing_enable()\n",
        "model = prepare_model_for_kbit_training(model)"
      ],
      "metadata": {
        "id": "a9EUEDAl0ss3"
      },
      "execution_count": null,
      "outputs": []
    },
    {
      "cell_type": "code",
      "source": [
        "def print_trainable_parameters(model):\n",
        "    \"\"\"\n",
        "    Prints the number of trainable parameters in the model.\n",
        "    \"\"\"\n",
        "    trainable_params = 0\n",
        "    all_param = 0\n",
        "    for _, param in model.named_parameters():\n",
        "        all_param += param.numel()\n",
        "        if param.requires_grad:\n",
        "            trainable_params += param.numel()\n",
        "    print(\n",
        "        f\"trainable params: {trainable_params} || all params: {all_param} || trainable%: {100 * trainable_params / all_param}\"\n",
        "    )"
      ],
      "metadata": {
        "id": "gkIcwsSU01EB"
      },
      "execution_count": null,
      "outputs": []
    },
    {
      "cell_type": "code",
      "source": [
        "from peft import LoraConfig, get_peft_model\n",
        "\n",
        "config = LoraConfig(\n",
        "    r=8, \n",
        "    lora_alpha=32, \n",
        "    target_modules=[\"query_key_value\"], \n",
        "    lora_dropout=0.05, \n",
        "    bias=\"none\", \n",
        "    task_type=\"CAUSAL_LM\"\n",
        ")\n",
        "\n",
        "model = get_peft_model(model, config)\n",
        "print_trainable_parameters(model)"
      ],
      "metadata": {
        "id": "Ybeyl20n3dYH"
      },
      "execution_count": null,
      "outputs": []
    },
    {
      "cell_type": "markdown",
      "source": [
        "# 데이터 학습하기"
      ],
      "metadata": {
        "id": "0ddGcswMUD9C"
      }
    },
    {
      "cell_type": "code",
      "source": [
        "import transformers\n",
        "\n",
        "# needed for gpt-neo-x tokenizer\n",
        "tokenizer.pad_token = tokenizer.eos_token\n",
        "\n",
        "trainer = transformers.Trainer(\n",
        "    model=model,\n",
        "    train_dataset=data[\"train\"],\n",
        "    args=transformers.TrainingArguments(\n",
        "        per_device_train_batch_size=2,\n",
        "        gradient_accumulation_steps=1,\n",
        "        max_steps=50, ## 초소량만 학습: 50 step만 학습. 약 4분정도 걸립니다.\n",
        "        learning_rate=1e-4,\n",
        "        fp16=True,\n",
        "        logging_steps=10,\n",
        "        output_dir=\"outputs\",\n",
        "        optim=\"paged_adamw_8bit\"\n",
        "    ),\n",
        "    data_collator=transformers.DataCollatorForLanguageModeling(tokenizer, mlm=False),\n",
        ")\n",
        "model.config.use_cache = False  # silence the warnings. Please re-enable for inference!\n",
        "trainer.train()"
      ],
      "metadata": {
        "id": "jq0nX33BmfaC"
      },
      "execution_count": null,
      "outputs": []
    },
    {
      "cell_type": "markdown",
      "source": [
        "# 모델 평가"
      ],
      "metadata": {
        "id": "LiM298jQUI5c"
      }
    },
    {
      "cell_type": "code",
      "source": [
        "model.eval()\n",
        "model.config.use_cache = True  # silence the warnings. Please re-enable for inference!"
      ],
      "metadata": {
        "id": "a-jauOEv9XVe"
      },
      "execution_count": null,
      "outputs": []
    },
    {
      "cell_type": "code",
      "source": [
        "def gen(x):\n",
        "    gened = model.generate(\n",
        "        **tokenizer(\n",
        "            f\"### 질문: {x}\\n\\n### 답변:\", \n",
        "            return_tensors='pt', \n",
        "            return_token_type_ids=False\n",
        "        ), \n",
        "        max_new_tokens=256,\n",
        "        early_stopping=True,\n",
        "        do_sample=True,\n",
        "        eos_token_id=2,\n",
        "    )\n",
        "    print(tokenizer.decode(gened[0]))"
      ],
      "metadata": {
        "id": "oDp9W-Gmp5Mb"
      },
      "execution_count": null,
      "outputs": []
    },
    {
      "cell_type": "markdown",
      "source": [
        "# 추론하기"
      ],
      "metadata": {
        "id": "08KlSlCoUN0X"
      }
    },
    {
      "cell_type": "code",
      "source": [
        "gen('카카오 간편로그인이 되지 않아요')"
      ],
      "metadata": {
        "id": "iIbK1GaipZd9"
      },
      "execution_count": null,
      "outputs": []
    },
    {
      "cell_type": "markdown",
      "source": [
        "# HuggingFaced에 모델 업로드"
      ],
      "metadata": {
        "id": "JriDpidBa3nG"
      }
    },
    {
      "cell_type": "code",
      "source": [
        "!huggingface-cli login"
      ],
      "metadata": {
        "id": "_LCQuL2na2ul"
      },
      "execution_count": null,
      "outputs": []
    },
    {
      "cell_type": "code",
      "source": [
        "model.push_to_hub('<ID>/<MODEL_NAME>')"
      ],
      "metadata": {
        "id": "TWc3sigabGiH"
      },
      "execution_count": null,
      "outputs": []
    }
  ]
}