{
  "nbformat": 4,
  "nbformat_minor": 0,
  "metadata": {
    "colab": {
      "provenance": [],
      "gpuType": "T4"
    },
    "kernelspec": {
      "name": "python3",
      "display_name": "Python 3"
    },
    "language_info": {
      "name": "python"
    },
    "accelerator": "GPU"
  },
  "cells": [
    {
      "cell_type": "markdown",
      "source": [
        "このnotebookは`stockmark/gpt-neox-japanese-1.4b`のモデルを`kunishou/databricks-dolly-15k-ja`のデータセットを用いてLoRA tuningするためのコードの例です。以下の例では、学習を1 epochを行います。T4 GPUで実行すると30分ほどかかります。\n",
        "\n",
        "- モデル：https://huggingface.co/stockmark/gpt-neox-japanese-1.4b\n",
        "- データ：https://github.com/kunishou/databricks-dolly-15k-ja\n",
        "\n",
        "\n",
        "また、ここで用いている設定は暫定的なもので、必要に応じて調整してください。"
      ],
      "metadata": {
        "id": "BPGgCZtMdMsv"
      }
    },
    {
      "cell_type": "markdown",
      "source": [
        "# ライブラリのインストール"
      ],
      "metadata": {
        "id": "hCZH9e6EcZyj"
      }
    },
    {
      "cell_type": "code",
      "execution_count": null,
      "metadata": {
        "id": "cmn52bx3v5Ha"
      },
      "outputs": [],
      "source": [
        "!python3 -m pip install -U pip\n",
        "!python3 -m pip install transformers accelerate datasets peft"
      ]
    },
    {
      "cell_type": "markdown",
      "source": [
        "# 準備"
      ],
      "metadata": {
        "id": "4t3Cqs9_ce3J"
      }
    },
    {
      "cell_type": "code",
      "source": [
        "import torch\n",
        "import datasets\n",
        "from transformers import AutoTokenizer, AutoModelForCausalLM, Trainer, TrainingArguments\n",
        "from peft import get_peft_model, LoraConfig, TaskType, PeftModel, PeftConfig\n",
        "\n",
        "model_name = \"stockmark/gpt-neox-japanese-1.4b\"\n",
        "peft_model_name = \"peft_model\"\n",
        "\n",
        "prompt_template = \"\"\"### Instruction:\n",
        "{instruction}\n",
        "\n",
        "### Input:\n",
        "{input}\n",
        "\n",
        "### Response:\n",
        "\"\"\"\n",
        "\n",
        "def encode(sample):\n",
        "    prompt = prompt_template.format(instruction=sample[\"instruction\"], input=sample[\"input\"])\n",
        "    target = sample[\"output\"] + tokenizer.eos_token\n",
        "    input_ids_prompt, input_ids_target = tokenizer([prompt, target]).input_ids\n",
        "    input_ids = input_ids_prompt + input_ids_target\n",
        "    labels = input_ids.copy()\n",
        "    labels[:len(input_ids_prompt)] = [-100] * len(input_ids_prompt)\n",
        "    return {\"input_ids\": input_ids, \"labels\": labels}\n",
        "\n",
        "def get_collator(tokenizer, max_length):\n",
        "    def collator(batch):\n",
        "        batch = [{ key: value[:max_length] for key, value in sample.items() } for sample in batch ]\n",
        "        batch = tokenizer.pad(batch, padding=True)\n",
        "        batch[\"labels\"] = [ e + [-100] * (len(batch[\"input_ids\"][0]) - len(e)) for e in batch[\"labels\"] ]\n",
        "        batch = { key: torch.tensor(value) for key, value in batch.items() }\n",
        "        return batch\n",
        "\n",
        "    return collator\n"
      ],
      "metadata": {
        "id": "hNdYMGMRzAVn"
      },
      "execution_count": null,
      "outputs": []
    },
    {
      "cell_type": "markdown",
      "source": [
        "# データセットとモデルの準備\n"
      ],
      "metadata": {
        "id": "UqXxPjJ_cliu"
      }
    },
    {
      "cell_type": "code",
      "source": [
        "# prepare dataset\n",
        "tokenizer = AutoTokenizer.from_pretrained(model_name)\n",
        "\n",
        "dataset_name = \"kunishou/databricks-dolly-15k-ja\"\n",
        "dataset = datasets.load_dataset(dataset_name)\n",
        "dataset = dataset.map(encode)\n",
        "dataset = dataset[\"train\"].train_test_split(0.2)\n",
        "train_dataset = dataset[\"train\"]\n",
        "val_dataset = dataset[\"test\"]\n",
        "\n",
        "# load model\n",
        "model = AutoModelForCausalLM.from_pretrained(model_name, device_map={\"\": 0}, torch_dtype=torch.float16)\n",
        "\n",
        "peft_config = LoraConfig(\n",
        "    task_type=TaskType.CAUSAL_LM,\n",
        "    inference_mode=False,\n",
        "    target_modules=[\"query_key_value\"],\n",
        "    r=16,\n",
        "    lora_alpha=32,\n",
        "    lora_dropout=0.05\n",
        ")\n",
        "\n",
        "model = get_peft_model(model, peft_config)\n",
        "model.print_trainable_parameters()"
      ],
      "metadata": {
        "id": "ZWdN-p7t0Grk"
      },
      "execution_count": null,
      "outputs": []
    },
    {
      "cell_type": "markdown",
      "source": [
        "# LoRA tuning"
      ],
      "metadata": {
        "id": "XCrdVAJYc88c"
      }
    },
    {
      "cell_type": "code",
      "source": [
        "training_args = TrainingArguments(\n",
        "    output_dir=\"./train_results\",\n",
        "    learning_rate=2e-4,\n",
        "    per_device_train_batch_size=4,\n",
        "    gradient_accumulation_steps=4,\n",
        "    per_device_eval_batch_size=16,\n",
        "    num_train_epochs=1,\n",
        "    logging_strategy='steps',\n",
        "    logging_steps=10,\n",
        "    save_strategy='epoch',\n",
        "    evaluation_strategy='epoch',\n",
        "    load_best_model_at_end=True,\n",
        "    metric_for_best_model=\"eval_loss\",\n",
        "    greater_is_better=False,\n",
        "    save_total_limit=2\n",
        ")\n",
        "\n",
        "trainer = Trainer(\n",
        "    model=model,\n",
        "    args=training_args,\n",
        "    train_dataset=train_dataset,\n",
        "    eval_dataset=val_dataset,\n",
        "    data_collator=get_collator(tokenizer, 512)\n",
        ")\n",
        "\n",
        "trainer.train()\n",
        "model = trainer.model\n",
        "model.save_pretrained(peft_model_name)"
      ],
      "metadata": {
        "id": "4LH9tOCTJVk1"
      },
      "execution_count": null,
      "outputs": []
    },
    {
      "cell_type": "markdown",
      "source": [
        "# 学習したモデルのロード"
      ],
      "metadata": {
        "id": "ORgzOPAqdEZR"
      }
    },
    {
      "cell_type": "code",
      "source": [
        "tokenizer = AutoTokenizer.from_pretrained(model_name)\n",
        "model = AutoModelForCausalLM.from_pretrained(model_name, device_map={\"\": 0}, torch_dtype=torch.float16)\n",
        "model = PeftModel.from_pretrained(model, peft_model_name)"
      ],
      "metadata": {
        "id": "yrExyO9EOvzR"
      },
      "execution_count": null,
      "outputs": []
    },
    {
      "cell_type": "markdown",
      "source": [
        "# 推論"
      ],
      "metadata": {
        "id": "-dttR6tkdG0k"
      }
    },
    {
      "cell_type": "code",
      "source": [
        "prompt = prompt_template.format(instruction=\"日本で人気のスポーツは？\", input=\"\")\n",
        "\n",
        "inputs = tokenizer(prompt, return_tensors=\"pt\").to(model.device)\n",
        "with torch.no_grad():\n",
        "    tokens = model.generate(\n",
        "        **inputs,\n",
        "        max_new_tokens=128,\n",
        "        repetition_penalty=1.1\n",
        "    )\n",
        "\n",
        "output = tokenizer.decode(tokens[0], skip_special_tokens=True)\n",
        "print(output)"
      ],
      "metadata": {
        "id": "pC5t9F1GJuFN"
      },
      "execution_count": null,
      "outputs": []
    }
  ]
}