ruti-v3.ipynb

ruti-v3.ipynb

{
  "nbformat": 4,
  "nbformat_minor": 0,
  "metadata": {
    "colab": {
      "provenance": [],
      "gpuType": "T4",
      "authorship_tag": "ABX9TyPjX1MQYl5a3LmhN4cgAP2c",
      "include_colab_link": true
    },
    "kernelspec": {
      "name": "python3",
      "display_name": "Python 3"
    },
    "language_info": {
      "name": "python"
    },
    "accelerator": "GPU"
  },
  "cells": [
    {
      "cell_type": "markdown",
      "metadata": {
        "id": "view-in-github",
        "colab_type": "text"
      },
      "source": [
        "<a href=\"https://colab.research.google.com/gist/chottokun/1530a48adc15bdb5190b08569bfd4bbd/ruti-v3.ipynb\" target=\"_parent\"><img src=\"https://colab.research.google.com/assets/colab-badge.svg\" alt=\"Open In Colab\"/></a>"
      ]
    },
    {
      "cell_type": "code",
      "execution_count": null,
      "metadata": {
        "id": "rhmttsjLpuNq"
      },
      "outputs": [],
      "source": [
        "!pip install -U \"transformers>=4.48.0\"\n",
        "# !pip install flash-attn --no-build-isolation"
      ]
    },
    {
      "cell_type": "code",
      "source": [
        "# テストコードを実行してみる。"
      ],
      "metadata": {
        "id": "ZF1kdF9wvaFS"
      },
      "execution_count": null,
      "outputs": []
    },
    {
      "cell_type": "code",
      "source": [
        "import torch.nn.functional as F\n",
        "from sentence_transformers import SentenceTransformer\n",
        "\n",
        "# Download from the 🤗 Hub\n",
        "model = SentenceTransformer(\"cl-nagoya/ruri-v3-310m\")\n",
        "\n",
        "# Ruri v3 employs a 1+3 prefix scheme to distinguish between different types of text inputs:\n",
        "# \"\" (empty string) is used for encoding semantic meaning.\n",
        "# \"トピック: \" is used for classification, clustering, and encoding topical information.\n",
        "# \"検索クエリ: \" is used for queries in retrieval tasks.\n",
        "# \"検索文書: \" is used for documents to be retrieved.\n",
        "sentences = [\n",
        "    \"川べりでサーフボードを持った人たちがいます\",\n",
        "    \"サーファーたちが川べりに立っています\",\n",
        "    \"トピック: 瑠璃色のサーファー\",\n",
        "    \"検索クエリ: 瑠璃色はどんな色？\",\n",
        "    \"検索文書: 瑠璃色（るりいろ）は、紫みを帯びた濃い青。名は、半貴石の瑠璃（ラピスラズリ、英: lapis lazuli）による。JIS慣用色名では「こい紫みの青」（略号 dp-pB）と定義している[1][2]。\",\n",
        "]\n",
        "\n",
        "embeddings = model.encode(sentences, convert_to_tensor=True)\n",
        "print(embeddings.size())\n",
        "# [5, 768]\n",
        "\n",
        "similarities = F.cosine_similarity(embeddings.unsqueeze(0), embeddings.unsqueeze(1), dim=2)\n",
        "print(similarities)\n",
        "# [[1.0000, 0.9603, 0.8157, 0.7074, 0.6916],\n",
        "#  [0.9603, 1.0000, 0.8192, 0.7014, 0.6819],\n",
        "#  [0.8157, 0.8192, 1.0000, 0.8701, 0.8470],\n",
        "#  [0.7074, 0.7014, 0.8701, 1.0000, 0.9746],\n",
        "#  [0.6916, 0.6819, 0.8470, 0.9746, 1.0000]]\n"
      ],
      "metadata": {
        "id": "xE-8pexgp7TY"
      },
      "execution_count": null,
      "outputs": []
    },
    {
      "cell_type": "markdown",
      "source": [
        "# langchainと使ってみる"
      ],
      "metadata": {
        "id": "baB9DuA_vJzA"
      }
    },
    {
      "cell_type": "code",
      "source": [
        "!pip install -q -U langchain-huggingface"
      ],
      "metadata": {
        "id": "taggKdYdqEr-"
      },
      "execution_count": null,
      "outputs": []
    },
    {
      "cell_type": "code",
      "source": [
        "from langchain_huggingface import HuggingFaceEmbeddings\n",
        "\n",
        "model_name = \"cl-nagoya/ruri-v3-310m\"\n",
        "model_kwargs = {'device': 'cuda'}\n",
        "encode_kwargs = {'normalize_embeddings': False}\n",
        "hf = HuggingFaceEmbeddings(\n",
        "    model_name=model_name,\n",
        "    model_kwargs=model_kwargs,\n",
        "    encode_kwargs=encode_kwargs\n",
        ")"
      ],
      "metadata": {
        "id": "YbH2Su9vqOQ_"
      },
      "execution_count": null,
      "outputs": []
    },
    {
      "cell_type": "code",
      "source": [
        "# テスト用の文章\n",
        "# テスト用の文章 (例)\n",
        "texts = [\n",
        "    \"今日は良い天気です。\",\n",
        "    \"明日は雨が降るでしょう。\",\n",
        "    \"週末は旅行に行きます。\",\n",
        "    \"美味しい料理を食べたいです。\",\n",
        "    \"新しい本を読んでいます。\",\n",
        "    \"映画を見に行きました。\",\n",
        "    \"音楽を聴いています。\",\n",
        "    \"ゲームで遊んでいます。\",\n",
        "    \"仕事で疲れています。\",\n",
        "    \"ゆっくり休みたいと思います。\",\n",
        "    \"友達と会いました。\",\n",
        "    \"家族と過ごしました。\",\n",
        "    \"誕生日プレゼントをもらいました。\",\n",
        "    \"クリスマスが楽しみです。\",\n",
        "    \"お正月は実家に帰ります。\",\n",
        "    \"明後日は嵐になります\",\n",
        "    \"桜が咲きました。\",\n",
        "    \"紅葉が綺麗です。\",\n",
        "    \"雪が降りました。\",\n",
        "    \"海に行きました。\",\n",
        "    \"山に登りました。\"\n",
        "]\n",
        "\n",
        "# 文章を埋め込みベクトルに変換\n",
        "embeddings = hf.embed_documents(texts)\n",
        "\n",
        "# 埋め込みベクトルのサイズを確認\n",
        "print(len(embeddings))  # 2\n",
        "print(len(embeddings[0]))  # 768\n",
        "\n",
        "# 埋め込みベクトルを表示 (最初の5次元のみ)\n",
        "print(embeddings[0][:5])\n",
        "print(embeddings[1][:5])"
      ],
      "metadata": {
        "id": "pw5uBu8sqfrP"
      },
      "execution_count": null,
      "outputs": []
    },
    {
      "cell_type": "code",
      "source": [
        "import numpy as np\n",
        "from sklearn.metrics.pairwise import cosine_similarity\n",
        "\n",
        "# 類似度を計算\n",
        "similarity_matrix = cosine_similarity(embeddings)\n",
        "\n",
        "# 類似度に基づいて文章を並べ替え\n",
        "for i in range(len(texts)):\n",
        "    # 各文章に対して、類似度が上位の文章を抽出\n",
        "    similar_indices = np.argsort(similarity_matrix[i])[::-1]\n",
        "\n",
        "    # 類似度と文章をまとめて表示\n",
        "    print(f\"--- {texts[i]} ---\")\n",
        "    for j in similar_indices:\n",
        "        if i != j:  # 自分自身を除外\n",
        "            print(f\"- {texts[j]} (類似度: {similarity_matrix[i][j]:.4f})\")"
      ],
      "metadata": {
        "id": "VWzvnEkEsdTc",
        "collapsed": true
      },
      "execution_count": null,
      "outputs": []
    },
    {
      "cell_type": "markdown",
      "source": [
        "# Retrieverとして"
      ],
      "metadata": {
        "id": "jMGbM88sth_f"
      }
    },
    {
      "cell_type": "code",
      "source": [
        "!pip install -q -U langchain-community\n",
        "!!pip install -q -U chromadb huggingface_hub"
      ],
      "metadata": {
        "id": "T287EDTctofn"
      },
      "execution_count": null,
      "outputs": []
    },
    {
      "cell_type": "code",
      "source": [
        "from langchain.embeddings import HuggingFaceEmbeddings\n",
        "from langchain.vectorstores import Chroma\n",
        "from langchain.chains import RetrievalQA\n",
        "from langchain.llms import HuggingFaceHub\n",
        "\n",
        "# 1. 埋め込みモデルとベクトルストアを初期化\n",
        "embeddings = HuggingFaceEmbeddings(\n",
        "    model_name=\"cl-nagoya/ruri-v3-310m\",\n",
        "    model_kwargs={\"device\": \"cuda\"},  # GPUを使用する場合\n",
        "    encode_kwargs={\"normalize_embeddings\": False},\n",
        ")\n",
        "# ChromaDBを使用\n",
        "db = Chroma.from_texts(texts, embeddings, persist_directory=\"db\")  # textsはテスト用の文章リスト、dbディレクトリに保存\n",
        "\n",
        "# 2. Retrieverを取得\n",
        "retriever = db.as_retriever()\n",
        "\n",
        "# 3. クエリを実行し、検索結果を取得\n",
        "query = \"今日の天気はどうですか？\"\n",
        "docs = retriever.get_relevant_documents(query)\n",
        "\n",
        "# 4. 検索結果を表示\n",
        "for doc in docs:\n",
        "    print(doc.page_content)\n",
        "    print(\"---\")"
      ],
      "metadata": {
        "id": "xrvTEYDztiX2"
      },
      "execution_count": null,
      "outputs": []
    }
  ]
}