{
 "cells": [
  {
   "cell_type": "markdown",
   "id": "41ca2df0",
   "metadata": {},
   "source": [
    "# 🤗 + 📚 dbmdz BERT and ELECTRA models\n"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "58e60a32",
   "metadata": {},
   "source": [
    "In this repository the MDZ Digital Library team (dbmdz) at the Bavarian State\n",
    "Library open sources Italian BERT and ELECTRA models 🎉\n"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "c7b5f379",
   "metadata": {},
   "source": [
    "# Italian BERT\n"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "5bf65013",
   "metadata": {},
   "source": [
    "The source data for the Italian BERT model consists of a recent Wikipedia dump and\n",
    "various texts from the [OPUS corpora](http://opus.nlpl.eu/) collection. The final\n",
    "training corpus has a size of 13GB and 2,050,057,573 tokens.\n"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "a3aadc8d",
   "metadata": {},
   "source": [
    "For sentence splitting, we use NLTK (faster compared to spacy).\n",
    "Our cased and uncased models are training with an initial sequence length of 512\n",
    "subwords for ~2-3M steps.\n"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "4d670485",
   "metadata": {},
   "source": [
    "For the XXL Italian models, we use the same training data from OPUS and extend\n",
    "it with data from the Italian part of the [OSCAR corpus](https://traces1.inria.fr/oscar/).\n",
    "Thus, the final training corpus has a size of 81GB and 13,138,379,147 tokens.\n"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "a366dc7d",
   "metadata": {},
   "source": [
    "Note: Unfortunately, a wrong vocab size was used when training the XXL models.\n",
    "This explains the mismatch of the \"real\" vocab size of 31102, compared to the\n",
    "vocab size specified in `config.json`. However, the model is working and all\n",
    "evaluations were done under those circumstances.\n",
    "See [this issue](https://github.com/dbmdz/berts/issues/7) for more information.\n"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "eaee3adf",
   "metadata": {},
   "source": [
    "The Italian ELECTRA model was trained on the \"XXL\" corpus for 1M steps in total using a batch\n",
    "size of 128. We pretty much following the ELECTRA training procedure as used for\n",
    "[BERTurk](https://github.com/stefan-it/turkish-bert/tree/master/electra).\n"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "c5151f8a",
   "metadata": {},
   "source": [
    "## Usage\n"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "id": "f9e48e19",
   "metadata": {},
   "outputs": [],
   "source": [
    "!pip install --upgrade paddlenlp"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "id": "c02e6f47",
   "metadata": {},
   "outputs": [],
   "source": [
    "import paddle\n",
    "from paddlenlp.transformers import AutoModel\n",
    "\n",
    "model = AutoModel.from_pretrained(\"dbmdz/bert-base-italian-xxl-cased\")\n",
    "input_ids = paddle.randint(100, 200, shape=[1, 20])\n",
    "print(model(input_ids))"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "13e03e4b",
   "metadata": {},
   "source": [
    "# Reference"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "66705724",
   "metadata": {},
   "source": [
    "> The model introduction and model weights originate from [https://huggingface.co/dbmdz/bert-base-italian-xxl-cased](https://huggingface.co/dbmdz/bert-base-italian-xxl-cased) and were converted to PaddlePaddle format for ease of use in PaddleNLP.\n"
   ]
  }
 ],
 "metadata": {
  "kernelspec": {
   "display_name": "Python 3 (ipykernel)",
   "language": "python",
   "name": "python3"
  },
  "language_info": {
   "codemirror_mode": {
    "name": "ipython",
    "version": 3
   },
   "file_extension": ".py",
   "mimetype": "text/x-python",
   "name": "python",
   "nbconvert_exporter": "python",
   "pygments_lexer": "ipython3",
   "version": "3.7.13"
  }
 },
 "nbformat": 4,
 "nbformat_minor": 5
}