diff --git a/Konopka_QA.ipynb b/Konopka_QA.ipynb
new file mode 100644
index 0000000..7836824
--- /dev/null
+++ b/Konopka_QA.ipynb
@@ -0,0 +1,2332 @@
+{
+ "cells": [
+  {
+   "cell_type": "markdown",
+   "source": [
+    "# Importy i sprawdzenie GPU"
+   ],
+   "metadata": {
+    "collapsed": false,
+    "pycharm": {
+     "name": "#%% md\n"
+    }
+   }
+  },
+  {
+   "cell_type": "code",
+   "execution_count": 1,
+   "outputs": [
+    {
+     "name": "stdout",
+     "output_type": "stream",
+     "text": [
+      "Available GPU True\n"
+     ]
+    }
+   ],
+   "source": [
+    "import datasets\n",
+    "import torch\n",
+    "import matplotlib.pyplot as plt\n",
+    "import pandas as pd\n",
+    "from transformers import AutoModelForQuestionAnswering, AutoTokenizer, pipeline\n",
+    "import json\n",
+    "import torch\n",
+    "\n",
+    "is_gpu_available = torch.cuda.is_available()\n",
+    "print(\"Available GPU\", is_gpu_available)"
+   ],
+   "metadata": {
+    "collapsed": false,
+    "pycharm": {
+     "name": "#%%\n"
+    }
+   }
+  },
+  {
+   "cell_type": "markdown",
+   "source": [
+    "# The dataset"
+   ],
+   "metadata": {
+    "collapsed": false,
+    "pycharm": {
+     "name": "#%% md\n"
+    }
+   }
+  },
+  {
+   "cell_type": "markdown",
+   "source": [
+    "Check SubjQA dataset domains"
+   ],
+   "metadata": {
+    "collapsed": false,
+    "pycharm": {
+     "name": "#%% md\n"
+    }
+   }
+  },
+  {
+   "cell_type": "code",
+   "execution_count": 2,
+   "metadata": {
+    "collapsed": true,
+    "pycharm": {
+     "name": "#%%\n"
+    }
+   },
+   "outputs": [
+    {
+     "data": {
+      "text/plain": "['books', 'electronics', 'grocery', 'movies', 'restaurants', 'tripadvisor']"
+     },
+     "execution_count": 2,
+     "metadata": {},
+     "output_type": "execute_result"
+    }
+   ],
+   "source": [
+    "datasets.get_dataset_config_names(\"subjqa\")"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": 3,
+   "outputs": [
+    {
+     "name": "stdout",
+     "output_type": "stream",
+     "text": [
+      "SubjQA is a question answering dataset that focuses on subjective questions and answers.\n",
+      "The dataset consists of roughly 10,000 questions over reviews from 6 different domains: books, movies, grocery,\n",
+      "electronics, TripAdvisor (i.e. hotels), and restaurants. \n",
+      "\n",
+      "FEATURES {'domain': Value(dtype='string', id=None), 'nn_mod': Value(dtype='string', id=None), 'nn_asp': Value(dtype='string', id=None), 'query_mod': Value(dtype='string', id=None), 'query_asp': Value(dtype='string', id=None), 'q_reviews_id': Value(dtype='string', id=None), 'question_subj_level': Value(dtype='int64', id=None), 'ques_subj_score': Value(dtype='float32', id=None), 'is_ques_subjective': Value(dtype='bool', id=None), 'review_id': Value(dtype='string', id=None), 'id': Value(dtype='string', id=None), 'title': Value(dtype='string', id=None), 'context': Value(dtype='string', id=None), 'question': Value(dtype='string', id=None), 'answers': Sequence(feature={'text': Value(dtype='string', id=None), 'answer_start': Value(dtype='int32', id=None), 'answer_subj_level': Value(dtype='int64', id=None), 'ans_subj_score': Value(dtype='float32', id=None), 'is_ans_subjective': Value(dtype='bool', id=None)}, length=-1, id=None)} \n",
+      "\n",
+      "SPLITS {'train': SplitInfo(name='train', num_bytes=1574953, num_examples=1165, dataset_name='subjqa'), 'test': SplitInfo(name='test', num_bytes=689440, num_examples=512, dataset_name='subjqa'), 'validation': SplitInfo(name='validation', num_bytes=312577, num_examples=230, dataset_name='subjqa')} \n",
+      "\n"
+     ]
+    }
+   ],
+   "source": [
+    "selected_config = 'tripadvisor'\n",
+    "ds_builder = datasets.load_dataset_builder(\"subjqa\", selected_config)\n",
+    "print(ds_builder.info.description, \"\\n\")\n",
+    "print(\"FEATURES\", ds_builder.info.features, \"\\n\")\n",
+    "print(\"SPLITS\", ds_builder.info.splits, \"\\n\")"
+   ],
+   "metadata": {
+    "collapsed": false,
+    "pycharm": {
+     "name": "#%%\n"
+    }
+   }
+  },
+  {
+   "cell_type": "code",
+   "execution_count": 4,
+   "outputs": [
+    {
+     "name": "stderr",
+     "output_type": "stream",
+     "text": [
+      "Reusing dataset subjqa (/home/karo/.cache/huggingface/datasets/subjqa/tripadvisor/1.1.0/e5588f9298ff2d70686a00cc377e4bdccf4e32287459e3c6baf2dc5ab57fe7fd)\n"
+     ]
+    },
+    {
+     "data": {
+      "text/plain": "  0%|          | 0/3 [00:00<?, ?it/s]",
+      "application/vnd.jupyter.widget-view+json": {
+       "version_major": 2,
+       "version_minor": 0,
+       "model_id": "232aa10fe30e48ccb8cdc7eaea268a23"
+      },
+      "application/json": {
+       "n": 0,
+       "total": 3,
+       "elapsed": 0.010135412216186523,
+       "ncols": null,
+       "nrows": 12,
+       "prefix": "",
+       "ascii": false,
+       "unit": "it",
+       "unit_scale": false,
+       "rate": null,
+       "bar_format": null,
+       "postfix": null,
+       "unit_divisor": 1000,
+       "initial": 0,
+       "colour": null
+      }
+     },
+     "metadata": {},
+     "output_type": "display_data"
+    }
+   ],
+   "source": [
+    "subjqa_tripadv = datasets.load_dataset(\"subjqa\", selected_config)"
+   ],
+   "metadata": {
+    "collapsed": false,
+    "pycharm": {
+     "name": "#%%\n"
+    }
+   }
+  },
+  {
+   "cell_type": "code",
+   "execution_count": 5,
+   "outputs": [
+    {
+     "name": "stdout",
+     "output_type": "stream",
+     "text": [
+      "SubjQA is a question answering dataset that focuses on subjective questions and answers.\n",
+      "The dataset consists of roughly 10,000 questions over reviews from 6 different domains: books, movies, grocery,\n",
+      "electronics, TripAdvisor (i.e. hotels), and restaurants. \n",
+      "\n",
+      "FEATURES {'domain': Value(dtype='string', id=None), 'nn_mod': Value(dtype='string', id=None), 'nn_asp': Value(dtype='string', id=None), 'query_mod': Value(dtype='string', id=None), 'query_asp': Value(dtype='string', id=None), 'q_reviews_id': Value(dtype='string', id=None), 'question_subj_level': Value(dtype='int64', id=None), 'ques_subj_score': Value(dtype='float32', id=None), 'is_ques_subjective': Value(dtype='bool', id=None), 'review_id': Value(dtype='string', id=None), 'id': Value(dtype='string', id=None), 'title': Value(dtype='string', id=None), 'context': Value(dtype='string', id=None), 'question': Value(dtype='string', id=None), 'answers': Sequence(feature={'text': Value(dtype='string', id=None), 'answer_start': Value(dtype='int32', id=None), 'answer_subj_level': Value(dtype='int64', id=None), 'ans_subj_score': Value(dtype='float32', id=None), 'is_ans_subjective': Value(dtype='bool', id=None)}, length=-1, id=None)} \n",
+      "\n",
+      "SPLITS {'train': SplitInfo(name='train', num_bytes=1823263, num_examples=1400, dataset_name='subjqa'), 'test': SplitInfo(name='test', num_bytes=335385, num_examples=266, dataset_name='subjqa'), 'validation': SplitInfo(name='validation', num_bytes=349286, num_examples=267, dataset_name='subjqa')} \n",
+      "\n"
+     ]
+    }
+   ],
+   "source": [
+    "selected_config = 'restaurants'\n",
+    "ds_builder = datasets.load_dataset_builder(\"subjqa\", selected_config)\n",
+    "print(ds_builder.info.description, \"\\n\")\n",
+    "print(\"FEATURES\", ds_builder.info.features, \"\\n\")\n",
+    "print(\"SPLITS\", ds_builder.info.splits, \"\\n\")"
+   ],
+   "metadata": {
+    "collapsed": false,
+    "pycharm": {
+     "name": "#%%\n"
+    }
+   }
+  },
+  {
+   "cell_type": "code",
+   "execution_count": 7,
+   "outputs": [
+    {
+     "name": "stderr",
+     "output_type": "stream",
+     "text": [
+      "Reusing dataset subjqa (/home/karo/.cache/huggingface/datasets/subjqa/restaurants/1.1.0/e5588f9298ff2d70686a00cc377e4bdccf4e32287459e3c6baf2dc5ab57fe7fd)\n"
+     ]
+    },
+    {
+     "data": {
+      "text/plain": "  0%|          | 0/3 [00:00<?, ?it/s]",
+      "application/vnd.jupyter.widget-view+json": {
+       "version_major": 2,
+       "version_minor": 0,
+       "model_id": "90f76784f16646c6ae91cdf888701a6c"
+      },
+      "application/json": {
+       "n": 0,
+       "total": 3,
+       "elapsed": 0.009796857833862305,
+       "ncols": null,
+       "nrows": 12,
+       "prefix": "",
+       "ascii": false,
+       "unit": "it",
+       "unit_scale": false,
+       "rate": null,
+       "bar_format": null,
+       "postfix": null,
+       "unit_divisor": 1000,
+       "initial": 0,
+       "colour": null
+      }
+     },
+     "metadata": {},
+     "output_type": "display_data"
+    }
+   ],
+   "source": [
+    "subjqa_restaurant = datasets.load_dataset(\"subjqa\", selected_config)"
+   ],
+   "metadata": {
+    "collapsed": false,
+    "pycharm": {
+     "name": "#%%\n"
+    }
+   }
+  },
+  {
+   "cell_type": "markdown",
+   "source": [
+    "Połącz datasety restaurant oraz trip_advisor oraz usuń test i przyłącz do train, ponieważ nie jest potrzebny"
+   ],
+   "metadata": {
+    "collapsed": false,
+    "pycharm": {
+     "name": "#%% md\n"
+    }
+   }
+  },
+  {
+   "cell_type": "code",
+   "execution_count": 8,
+   "outputs": [
+    {
+     "data": {
+      "text/plain": "DatasetDict({\n    train: Dataset({\n        features: ['domain', 'nn_mod', 'nn_asp', 'query_mod', 'query_asp', 'q_reviews_id', 'question_subj_level', 'ques_subj_score', 'is_ques_subjective', 'review_id', 'id', 'title', 'context', 'question', 'answers'],\n        num_rows: 3343\n    })\n    validation: Dataset({\n        features: ['domain', 'nn_mod', 'nn_asp', 'query_mod', 'query_asp', 'q_reviews_id', 'question_subj_level', 'ques_subj_score', 'is_ques_subjective', 'review_id', 'id', 'title', 'context', 'question', 'answers'],\n        num_rows: 497\n    })\n})"
+     },
+     "execution_count": 8,
+     "metadata": {},
+     "output_type": "execute_result"
+    }
+   ],
+   "source": [
+    "def concat_test2train(dataset):\n",
+    "    dataset['train'] = datasets.concatenate_datasets([dataset['test'], dataset['train']])\n",
+    "    del dataset['test']\n",
+    "    return dataset\n",
+    "\n",
+    "\n",
+    "subjqa_restaurant = concat_test2train(subjqa_restaurant)\n",
+    "subjqa_tripadv = concat_test2train(subjqa_tripadv)\n",
+    "subjqa = datasets.DatasetDict({\n",
+    "    'train': datasets.concatenate_datasets([subjqa_tripadv['train'], subjqa_restaurant['train']]),\n",
+    "    'validation': datasets.concatenate_datasets([subjqa_tripadv['validation'], subjqa_restaurant['validation']]),\n",
+    "})\n",
+    "\n",
+    "subjqa\n"
+   ],
+   "metadata": {
+    "collapsed": false,
+    "pycharm": {
+     "name": "#%%\n"
+    }
+   }
+  },
+  {
+   "cell_type": "code",
+   "execution_count": 10,
+   "outputs": [
+    {
+     "name": "stdout",
+     "output_type": "stream",
+     "text": [
+      "Is a cramped  room?\n",
+      "{'text': ['Hotel rooms were very small', 'rooms were very small'], 'answer_start': [12, 18], 'answer_subj_level': [2, 2], 'ans_subj_score': [0.5199999809265137, 0.5199999809265137], 'is_ans_subjective': [True, True]}\n"
+     ]
+    }
+   ],
+   "source": [
+    "print(subjqa[\"train\"]['question'][0])\n",
+    "print(subjqa[\"train\"][\"answers\"][0])"
+   ],
+   "metadata": {
+    "collapsed": false,
+    "pycharm": {
+     "name": "#%%\n"
+    }
+   }
+  },
+  {
+   "cell_type": "markdown",
+   "source": [
+    "Ilość unikalnych pytań w split"
+   ],
+   "metadata": {
+    "collapsed": false,
+    "pycharm": {
+     "name": "#%% md\n"
+    }
+   }
+  },
+  {
+   "cell_type": "code",
+   "execution_count": 11,
+   "outputs": [
+    {
+     "name": "stdout",
+     "output_type": "stream",
+     "text": [
+      "Number of questions in train: 3343\n",
+      "Number of questions in validation: 497\n"
+     ]
+    }
+   ],
+   "source": [
+    "dfs = {split: dset.to_pandas() for split, dset in subjqa.flatten().items()}\n",
+    "\n",
+    "for split, df in dfs.items():\n",
+    "    print(f\"Number of questions in {split}: {df['id'].nunique()}\")"
+   ],
+   "metadata": {
+    "collapsed": false,
+    "pycharm": {
+     "name": "#%%\n"
+    }
+   }
+  },
+  {
+   "cell_type": "code",
+   "execution_count": 12,
+   "outputs": [
+    {
+     "data": {
+      "text/plain": "                                    id                           title  \\\n2544  0d78196c65a3bec3913e2291cd9b771f          Yl2TN9c23ZGLUBSD9ks5Uw   \n884   57b688d836d9502a076fe6a50a9cde81  usa_san francisco_hotel_adagio   \n\n                                 question                       answers.text  \\\n2544                      Is it present ?  [so it's a bit of a splurge meal]   \n884   Where can I locate the hotel staff?                                 []   \n\n     answers.answer_start                                            context  \n2544                [177]  Byblos has a really beautiful decor, that I wo...  \n884                    []  Stayed at the Adagio for 4 nights end of June-...  ",
+      "text/html": "<div>\n<style scoped>\n    .dataframe tbody tr th:only-of-type {\n        vertical-align: middle;\n    }\n\n    .dataframe tbody tr th {\n        vertical-align: top;\n    }\n\n    .dataframe thead th {\n        text-align: right;\n    }\n</style>\n<table border=\"1\" class=\"dataframe\">\n  <thead>\n    <tr style=\"text-align: right;\">\n      <th></th>\n      <th>id</th>\n      <th>title</th>\n      <th>question</th>\n      <th>answers.text</th>\n      <th>answers.answer_start</th>\n      <th>context</th>\n    </tr>\n  </thead>\n  <tbody>\n    <tr>\n      <th>2544</th>\n      <td>0d78196c65a3bec3913e2291cd9b771f</td>\n      <td>Yl2TN9c23ZGLUBSD9ks5Uw</td>\n      <td>Is it present ?</td>\n      <td>[so it's a bit of a splurge meal]</td>\n      <td>[177]</td>\n      <td>Byblos has a really beautiful decor, that I wo...</td>\n    </tr>\n    <tr>\n      <th>884</th>\n      <td>57b688d836d9502a076fe6a50a9cde81</td>\n      <td>usa_san francisco_hotel_adagio</td>\n      <td>Where can I locate the hotel staff?</td>\n      <td>[]</td>\n      <td>[]</td>\n      <td>Stayed at the Adagio for 4 nights end of June-...</td>\n    </tr>\n  </tbody>\n</table>\n</div>"
+     },
+     "execution_count": 12,
+     "metadata": {},
+     "output_type": "execute_result"
+    }
+   ],
+   "source": [
+    "qa_cols = [\"id\", \"title\", \"question\", \"answers.text\",\n",
+    "           \"answers.answer_start\", \"context\"]\n",
+    "sample_df = dfs[\"train\"][qa_cols].sample(2, random_state=7)\n",
+    "sample_df"
+   ],
+   "metadata": {
+    "collapsed": false,
+    "pycharm": {
+     "name": "#%%\n"
+    }
+   }
+  },
+  {
+   "cell_type": "markdown",
+   "source": [
+    "Sprawdzenie generowania odpowiedzi z kontekstu"
+   ],
+   "metadata": {
+    "collapsed": false,
+    "pycharm": {
+     "name": "#%% md\n"
+    }
+   }
+  },
+  {
+   "cell_type": "code",
+   "execution_count": 13,
+   "outputs": [
+    {
+     "data": {
+      "text/plain": "\"so it's a bit of a splurge meal\""
+     },
+     "execution_count": 13,
+     "metadata": {},
+     "output_type": "execute_result"
+    }
+   ],
+   "source": [
+    "start_idx = sample_df[\"answers.answer_start\"].iloc[0][0]\n",
+    "end_idx = start_idx + len(sample_df[\"answers.text\"].iloc[0][0])\n",
+    "sample_df[\"context\"].iloc[0][start_idx:end_idx]"
+   ],
+   "metadata": {
+    "collapsed": false,
+    "pycharm": {
+     "name": "#%%\n"
+    }
+   }
+  },
+  {
+   "cell_type": "markdown",
+   "source": [
+    "Usuń obiekty które nie posiadają odpowiedzi"
+   ],
+   "metadata": {
+    "collapsed": false,
+    "pycharm": {
+     "name": "#%% md\n"
+    }
+   }
+  },
+  {
+   "cell_type": "code",
+   "execution_count": 16,
+   "outputs": [
+    {
+     "data": {
+      "text/plain": "  0%|          | 0/2 [00:00<?, ?ba/s]",
+      "application/vnd.jupyter.widget-view+json": {
+       "version_major": 2,
+       "version_minor": 0,
+       "model_id": "dc2ed0edd4b6473d9e419ca242b6a06d"
+      },
+      "application/json": {
+       "n": 0,
+       "total": 2,
+       "elapsed": 0.00982809066772461,
+       "ncols": null,
+       "nrows": 12,
+       "prefix": "",
+       "ascii": false,
+       "unit": "ba",
+       "unit_scale": false,
+       "rate": null,
+       "bar_format": null,
+       "postfix": null,
+       "unit_divisor": 1000,
+       "initial": 0,
+       "colour": null
+      }
+     },
+     "metadata": {},
+     "output_type": "display_data"
+    },
+    {
+     "data": {
+      "text/plain": "  0%|          | 0/1 [00:00<?, ?ba/s]",
+      "application/vnd.jupyter.widget-view+json": {
+       "version_major": 2,
+       "version_minor": 0,
+       "model_id": "4712d554c86d49d491be1c141f45bafd"
+      },
+      "application/json": {
+       "n": 0,
+       "total": 1,
+       "elapsed": 0.009340763092041016,
+       "ncols": null,
+       "nrows": 12,
+       "prefix": "",
+       "ascii": false,
+       "unit": "ba",
+       "unit_scale": false,
+       "rate": null,
+       "bar_format": null,
+       "postfix": null,
+       "unit_divisor": 1000,
+       "initial": 0,
+       "colour": null
+      }
+     },
+     "metadata": {},
+     "output_type": "display_data"
+    },
+    {
+     "data": {
+      "text/plain": "DatasetDict({\n    train: Dataset({\n        features: ['domain', 'nn_mod', 'nn_asp', 'query_mod', 'query_asp', 'q_reviews_id', 'question_subj_level', 'ques_subj_score', 'is_ques_subjective', 'review_id', 'id', 'title', 'context', 'question', 'answers'],\n        num_rows: 1666\n    })\n    validation: Dataset({\n        features: ['domain', 'nn_mod', 'nn_asp', 'query_mod', 'query_asp', 'q_reviews_id', 'question_subj_level', 'ques_subj_score', 'is_ques_subjective', 'review_id', 'id', 'title', 'context', 'question', 'answers'],\n        num_rows: 265\n    })\n})"
+     },
+     "execution_count": 16,
+     "metadata": {},
+     "output_type": "execute_result"
+    }
+   ],
+   "source": [
+    "subjqa['train'] = subjqa['train'].filter(lambda example: len(example[\"answers\"]['text']) > 0)\n",
+    "subjqa['validation'] = subjqa['validation'].filter(lambda example: len(example[\"answers\"]['text']) > 0)\n",
+    "dfs = {split: dset.to_pandas() for split, dset in subjqa.flatten().items()}\n",
+    "subjqa"
+   ],
+   "metadata": {
+    "collapsed": false,
+    "pycharm": {
+     "name": "#%%\n"
+    }
+   }
+  },
+  {
+   "cell_type": "code",
+   "execution_count": 15,
+   "outputs": [
+    {
+     "data": {
+      "text/plain": "        domain      nn_mod           nn_asp query_mod        query_asp  \\\n0  tripadvisor   excellent            hotel  friendly            hotel   \n1  tripadvisor        neat            hotel      cozy            hotel   \n2  tripadvisor   excellent  value for money      good  value for money   \n3  tripadvisor  convenient            place      safe            hotel   \n4  tripadvisor     helpful            staff  helpfull            staff   \n\n                       q_reviews_id  question_subj_level  ques_subj_score  \\\n0  b96b8478f5202ac9534eaf75167016f7                    1              0.0   \n1  ca174c824baba906d30eed207350e37e                    1              0.0   \n2  5d41bfd0e2166e14eeae1c1be9085555                    3              0.0   \n3  a5880d95aa1161cfaac48584ee58934d                    1              0.6   \n4  902b8e2a5c10f1796abdc830c4a4acd2                    2              0.0   \n\n   is_ques_subjective                review_id  \\\n0               False  tripadvisor_review_1509   \n1               False  tripadvisor_review_6133   \n2               False  tripadvisor_review_4872   \n3                True  tripadvisor_review_4715   \n4               False  tripadvisor_review_1093   \n\n                                 id  \\\n0  d1c352b70d1225245569a0a1acbf5e04   \n1  b6fbf58ca273ad8f9b47a9be6a36e707   \n2  ea95bccdd762284ad7040be8d016da4f   \n3  413df1095d03e4f967d349f0490a2514   \n4  d7fdc86b464f2a797dca33b058b68078   \n\n                                               title  \\\n0   usa_san francisco_argonaut_hotel_a_kimpton_hotel   \n1  usa_san francisco_best_western_tuscan_inn_fish...   \n2                       usa_san francisco_castle_inn   \n3                       usa_san francisco_castle_inn   \n4  usa_san francisco_chancellor_hotel_on_union_sq...   \n\n                                             context  \\\n0  Great setting at the end of the wharf (so your...   \n1  My wife and I took two trips to San Fran in 20...   \n2  Yep, I have to agree with all those folks who ...   \n3  On first sight the Castle hotel is not great, ...   \n4  Stayed at the Chancellor recently for 3 nights...   \n\n                             question  \\\n0                  How was the hotel?   \n1                   How is the hotel?   \n2              Is it value for money?   \n3  Does the hotel offer good service?   \n4          How do you rate the staff?   \n\n                                        answers.text answers.answer_start  \\\n0                                 [excellent hotels]                [527]   \n1                     [The hotel location was great]                [129]   \n2  [And very reasonably priced. Overall, excellen...                [501]   \n3     [On first sight the Castle hotel is not great]                  [0]   \n4                               [Staff very helpful]                 [86]   \n\n  answers.answer_subj_level answers.ans_subj_score answers.is_ans_subjective  \n0                       [1]                  [1.0]                    [True]  \n1                       [1]                 [0.75]                    [True]  \n2                       [3]                 [0.58]                    [True]  \n3                       [1]            [0.5416667]                    [True]  \n4                       [2]                  [0.3]                   [False]  ",
+      "text/html": "<div>\n<style scoped>\n    .dataframe tbody tr th:only-of-type {\n        vertical-align: middle;\n    }\n\n    .dataframe tbody tr th {\n        vertical-align: top;\n    }\n\n    .dataframe thead th {\n        text-align: right;\n    }\n</style>\n<table border=\"1\" class=\"dataframe\">\n  <thead>\n    <tr style=\"text-align: right;\">\n      <th></th>\n      <th>domain</th>\n      <th>nn_mod</th>\n      <th>nn_asp</th>\n      <th>query_mod</th>\n      <th>query_asp</th>\n      <th>q_reviews_id</th>\n      <th>question_subj_level</th>\n      <th>ques_subj_score</th>\n      <th>is_ques_subjective</th>\n      <th>review_id</th>\n      <th>id</th>\n      <th>title</th>\n      <th>context</th>\n      <th>question</th>\n      <th>answers.text</th>\n      <th>answers.answer_start</th>\n      <th>answers.answer_subj_level</th>\n      <th>answers.ans_subj_score</th>\n      <th>answers.is_ans_subjective</th>\n    </tr>\n  </thead>\n  <tbody>\n    <tr>\n      <th>0</th>\n      <td>tripadvisor</td>\n      <td>excellent</td>\n      <td>hotel</td>\n      <td>friendly</td>\n      <td>hotel</td>\n      <td>b96b8478f5202ac9534eaf75167016f7</td>\n      <td>1</td>\n      <td>0.0</td>\n      <td>False</td>\n      <td>tripadvisor_review_1509</td>\n      <td>d1c352b70d1225245569a0a1acbf5e04</td>\n      <td>usa_san francisco_argonaut_hotel_a_kimpton_hotel</td>\n      <td>Great setting at the end of the wharf (so your...</td>\n      <td>How was the hotel?</td>\n      <td>[excellent hotels]</td>\n      <td>[527]</td>\n      <td>[1]</td>\n      <td>[1.0]</td>\n      <td>[True]</td>\n    </tr>\n    <tr>\n      <th>1</th>\n      <td>tripadvisor</td>\n      <td>neat</td>\n      <td>hotel</td>\n      <td>cozy</td>\n      <td>hotel</td>\n      <td>ca174c824baba906d30eed207350e37e</td>\n      <td>1</td>\n      <td>0.0</td>\n      <td>False</td>\n      <td>tripadvisor_review_6133</td>\n      <td>b6fbf58ca273ad8f9b47a9be6a36e707</td>\n      <td>usa_san francisco_best_western_tuscan_inn_fish...</td>\n      <td>My wife and I took two trips to San Fran in 20...</td>\n      <td>How is the hotel?</td>\n      <td>[The hotel location was great]</td>\n      <td>[129]</td>\n      <td>[1]</td>\n      <td>[0.75]</td>\n      <td>[True]</td>\n    </tr>\n    <tr>\n      <th>2</th>\n      <td>tripadvisor</td>\n      <td>excellent</td>\n      <td>value for money</td>\n      <td>good</td>\n      <td>value for money</td>\n      <td>5d41bfd0e2166e14eeae1c1be9085555</td>\n      <td>3</td>\n      <td>0.0</td>\n      <td>False</td>\n      <td>tripadvisor_review_4872</td>\n      <td>ea95bccdd762284ad7040be8d016da4f</td>\n      <td>usa_san francisco_castle_inn</td>\n      <td>Yep, I have to agree with all those folks who ...</td>\n      <td>Is it value for money?</td>\n      <td>[And very reasonably priced. Overall, excellen...</td>\n      <td>[501]</td>\n      <td>[3]</td>\n      <td>[0.58]</td>\n      <td>[True]</td>\n    </tr>\n    <tr>\n      <th>3</th>\n      <td>tripadvisor</td>\n      <td>convenient</td>\n      <td>place</td>\n      <td>safe</td>\n      <td>hotel</td>\n      <td>a5880d95aa1161cfaac48584ee58934d</td>\n      <td>1</td>\n      <td>0.6</td>\n      <td>True</td>\n      <td>tripadvisor_review_4715</td>\n      <td>413df1095d03e4f967d349f0490a2514</td>\n      <td>usa_san francisco_castle_inn</td>\n      <td>On first sight the Castle hotel is not great, ...</td>\n      <td>Does the hotel offer good service?</td>\n      <td>[On first sight the Castle hotel is not great]</td>\n      <td>[0]</td>\n      <td>[1]</td>\n      <td>[0.5416667]</td>\n      <td>[True]</td>\n    </tr>\n    <tr>\n      <th>4</th>\n      <td>tripadvisor</td>\n      <td>helpful</td>\n      <td>staff</td>\n      <td>helpfull</td>\n      <td>staff</td>\n      <td>902b8e2a5c10f1796abdc830c4a4acd2</td>\n      <td>2</td>\n      <td>0.0</td>\n      <td>False</td>\n      <td>tripadvisor_review_1093</td>\n      <td>d7fdc86b464f2a797dca33b058b68078</td>\n      <td>usa_san francisco_chancellor_hotel_on_union_sq...</td>\n      <td>Stayed at the Chancellor recently for 3 nights...</td>\n      <td>How do you rate the staff?</td>\n      <td>[Staff very helpful]</td>\n      <td>[86]</td>\n      <td>[2]</td>\n      <td>[0.3]</td>\n      <td>[False]</td>\n    </tr>\n  </tbody>\n</table>\n</div>"
+     },
+     "execution_count": 15,
+     "metadata": {},
+     "output_type": "execute_result"
+    }
+   ],
+   "source": [
+    "dfs['validation'].head()"
+   ],
+   "metadata": {
+    "collapsed": false,
+    "pycharm": {
+     "name": "#%%\n"
+    }
+   }
+  },
+  {
+   "cell_type": "markdown",
+   "source": [
+    "Sprawdzenie typów pytań"
+   ],
+   "metadata": {
+    "collapsed": false,
+    "pycharm": {
+     "name": "#%% md\n"
+    }
+   }
+  },
+  {
+   "cell_type": "code",
+   "execution_count": 17,
+   "outputs": [
+    {
+     "data": {
+      "text/plain": "<Figure size 640x480 with 1 Axes>",
+      "image/png": "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\n"
+     },
+     "metadata": {},
+     "output_type": "display_data"
+    },
+    {
+     "data": {
+      "text/plain": "<Figure size 640x480 with 1 Axes>",
+      "image/png": "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\n"
+     },
+     "metadata": {},
+     "output_type": "display_data"
+    }
+   ],
+   "source": [
+    "def show_question_types_counts(dfs_split):\n",
+    "    counts = {}\n",
+    "    question_types = [\"What\", \"How\", \"Is\", \"Does\", \"Do\", \"Was\", \"Where\", \"Why\"]\n",
+    "\n",
+    "    for q in question_types:\n",
+    "        try:\n",
+    "            counts[q] = dfs[dfs_split][\"question\"].str.startswith(q).value_counts()[True]\n",
+    "        except:\n",
+    "            counts[q] = 0\n",
+    "    pd.Series(counts).sort_values().plot.barh()\n",
+    "    plt.title(\"Frequency of Question Types \" + dfs_split)\n",
+    "    plt.show()\n",
+    "\n",
+    "\n",
+    "show_question_types_counts(\"train\")\n",
+    "show_question_types_counts(\"validation\")"
+   ],
+   "metadata": {
+    "collapsed": false,
+    "pycharm": {
+     "name": "#%%\n"
+    }
+   }
+  },
+  {
+   "cell_type": "markdown",
+   "source": [
+    "Pytania i ilości odpowiedzi"
+   ],
+   "metadata": {
+    "collapsed": false,
+    "pycharm": {
+     "name": "#%% md\n"
+    }
+   }
+  },
+  {
+   "cell_type": "code",
+   "execution_count": 18,
+   "outputs": [
+    {
+     "data": {
+      "text/plain": "<Figure size 640x480 with 1 Axes>",
+      "image/png": "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\n"
+     },
+     "metadata": {},
+     "output_type": "display_data"
+    },
+    {
+     "data": {
+      "text/plain": "<Figure size 640x480 with 1 Axes>",
+      "image/png": "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\n"
+     },
+     "metadata": {},
+     "output_type": "display_data"
+    }
+   ],
+   "source": [
+    "def show_answers_counts(dfs_split):\n",
+    "    counts = dfs[dfs_split][\"answers.text\"] \\\n",
+    "        .groupby(dfs[dfs_split][\"answers.text\"].str.len()) \\\n",
+    "        .count().to_dict()\n",
+    "\n",
+    "    pd.Series(counts).sort_values().plot.barh()\n",
+    "    plt.title(\"Number of Answers \" + dfs_split)\n",
+    "    plt.show()\n",
+    "\n",
+    "\n",
+    "show_answers_counts(\"train\")\n",
+    "show_answers_counts(\"validation\")\n"
+   ],
+   "metadata": {
+    "collapsed": false,
+    "pycharm": {
+     "name": "#%%\n"
+    }
+   }
+  },
+  {
+   "cell_type": "markdown",
+   "source": [
+    "Przykładowe pytania"
+   ],
+   "metadata": {
+    "collapsed": false,
+    "pycharm": {
+     "name": "#%% md\n"
+    }
+   }
+  },
+  {
+   "cell_type": "code",
+   "execution_count": 19,
+   "outputs": [
+    {
+     "name": "stdout",
+     "output_type": "stream",
+     "text": [
+      "How was the service staff?\n",
+      "How is the stay?\n",
+      "How tastefully decorated was the room?\n",
+      "What is the customer service?\n",
+      "What do you think about dinner?\n",
+      "What is the most expensive price of food?\n",
+      "Is it service ?\n",
+      "Is it location ?\n",
+      "Is it a good place to stay?\n"
+     ]
+    }
+   ],
+   "source": [
+    "for question_type in [\"How\", \"What\", \"Is\"]:\n",
+    "    for question in (\n",
+    "            dfs[\"train\"][dfs[\"train\"].question.str.startswith(question_type)]\n",
+    "                    .sample(n=3, random_state=42)['question']):\n",
+    "        print(question)"
+   ],
+   "metadata": {
+    "collapsed": false,
+    "pycharm": {
+     "name": "#%%\n"
+    }
+   }
+  },
+  {
+   "cell_type": "markdown",
+   "source": [
+    "# Preprocessing"
+   ],
+   "metadata": {
+    "collapsed": false,
+    "pycharm": {
+     "name": "#%% md\n"
+    }
+   }
+  },
+  {
+   "cell_type": "markdown",
+   "source": [
+    "## Tokenizacja"
+   ],
+   "metadata": {
+    "collapsed": false,
+    "pycharm": {
+     "name": "#%% md\n"
+    }
+   }
+  },
+  {
+   "cell_type": "code",
+   "execution_count": 20,
+   "outputs": [
+    {
+     "data": {
+      "text/plain": "True"
+     },
+     "execution_count": 20,
+     "metadata": {},
+     "output_type": "execute_result"
+    }
+   ],
+   "source": [
+    "model_ckpt = \"bert-base-cased\"\n",
+    "tokenizer = AutoTokenizer.from_pretrained(model_ckpt)\n",
+    "tokenizer.is_fast"
+   ],
+   "metadata": {
+    "collapsed": false,
+    "pycharm": {
+     "name": "#%%\n"
+    }
+   }
+  },
+  {
+   "cell_type": "code",
+   "execution_count": 21,
+   "outputs": [
+    {
+     "name": "stderr",
+     "output_type": "stream",
+     "text": [
+      "Token indices sequence length is longer than the specified maximum sequence length for this model (556 > 512). Running this sequence through the model will result in indexing errors\n"
+     ]
+    },
+    {
+     "data": {
+      "text/plain": "<Figure size 640x480 with 1 Axes>",
+      "image/png": "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\n"
+     },
+     "metadata": {},
+     "output_type": "display_data"
+    }
+   ],
+   "source": [
+    "def compute_input_length(row):\n",
+    "    inputs = tokenizer(row[\"question\"], row[\"context\"])\n",
+    "    inputs_length = len(inputs[\"input_ids\"])\n",
+    "    del inputs\n",
+    "    return inputs_length\n",
+    "\n",
+    "\n",
+    "dfs[\"train\"][\"n_tokens\"] = dfs[\"train\"].apply(compute_input_length, axis=1)\n",
+    "\n",
+    "fig, ax = plt.subplots()\n",
+    "dfs[\"train\"][\"n_tokens\"].hist(bins=100, grid=False, ec=\"C0\", ax=ax)\n",
+    "plt.xlabel(\"Number of tokens in question-context pair\")\n",
+    "ax.axvline(x=384, ymin=0, ymax=1, linestyle=\"--\", color=\"C1\",\n",
+    "           label=\"Maximum sequence length\")\n",
+    "plt.legend()\n",
+    "plt.ylabel(\"Count\")\n",
+    "plt.show()"
+   ],
+   "metadata": {
+    "collapsed": false,
+    "pycharm": {
+     "name": "#%%\n"
+    }
+   }
+  },
+  {
+   "cell_type": "markdown",
+   "source": [
+    "Użycie stride w celu pozyskania okna z kontekstem i okrojenia wielkości tesktu, prezentacja działania"
+   ],
+   "metadata": {
+    "collapsed": false,
+    "pycharm": {
+     "name": "#%% md\n"
+    }
+   }
+  },
+  {
+   "cell_type": "code",
+   "execution_count": 22,
+   "outputs": [
+    {
+     "name": "stdout",
+     "output_type": "stream",
+     "text": [
+      "[CLS] How many employees does the company have? [SEP] We had a great two night stay in Dec 2008. Hotel staff were very accomodating with our 18 month old - providing a cot, toy and stroller for him. The staff were friendly and upbeat. Our room was also upgraded. Our only dissappointment was the food at the downstairs restaurant - way too expensive and did not meet expectations. [SEP]\n",
+      "[CLS] Was the atmosphere of the tourist areas of san fransisco peaceful? [SEP] My wife and I stayed here on our honeymoon for 2 nights in early March and had a great stay. I had booked a room at another hotel on hotels. com on Jan. 6th though. Then 4 days before we were to start our honeymoon in San Fran., hotels. com send me an e - mail to say they had to change my reservation from a hotel downtown to a hotel at the [SEP]\n",
+      "[CLS] Was the atmosphere of the tourist areas of san fransisco peaceful? [SEP]. com on Jan. 6th though. Then 4 days before we were to start our honeymoon in San Fran., hotels. com send me an e - mail to say they had to change my reservation from a hotel downtown to a hotel at the airport. I rufused and was given a refund and no hotel. There was a huge 30, 000 person convention in town at the same [SEP]\n",
+      "[CLS] Was the atmosphere of the tourist areas of san fransisco peaceful? [SEP] - mail to say they had to change my reservation from a hotel downtown to a hotel at the airport. I rufused and was given a refund and no hotel. There was a huge 30, 000 person convention in town at the same time so all the hotels were full. I got lucky though and was able to book a bay view room at The Argonaut. This was the best [SEP]\n",
+      "[CLS] Was the atmosphere of the tourist areas of san fransisco peaceful? [SEP]nd and no hotel. There was a huge 30, 000 person convention in town at the same time so all the hotels were full. I got lucky though and was able to book a bay view room at The Argonaut. This was the best thing that could of happened to us! This place was amazing! We arrived at the hotel around 12 : 30pm and were able to check in right away [SEP]\n",
+      "[CLS] Was the atmosphere of the tourist areas of san fransisco peaceful? [SEP] and was able to book a bay view room at The Argonaut. This was the best thing that could of happened to us! This place was amazing! We arrived at the hotel around 12 : 30pm and were able to check in right away! The staff were all so nice and called us by name whenever they saw us. The room was amazing! We had a beautiful view of Alcatraz [SEP]\n",
+      "[CLS] Was the atmosphere of the tourist areas of san fransisco peaceful? [SEP]! We arrived at the hotel around 12 : 30pm and were able to check in right away! The staff were all so nice and called us by name whenever they saw us. The room was amazing! We had a beautiful view of Alcatraz Island and could even see the golden gate. Internet worked great and the TV was fine. It's not a big fancy TV, but we didn'[SEP]\n",
+      "[CLS] Was the atmosphere of the tourist areas of san fransisco peaceful? [SEP] whenever they saw us. The room was amazing! We had a beautiful view of Alcatraz Island and could even see the golden gate. Internet worked great and the TV was fine. It's not a big fancy TV, but we didn't fly all the way to San Fran. to watch TV anyway. What made our stay so great was the concierge. There was a wonderful [SEP]\n",
+      "[CLS] Was the atmosphere of the tourist areas of san fransisco peaceful? [SEP] and the TV was fine. It's not a big fancy TV, but we didn't fly all the way to San Fran. to watch TV anyway. What made our stay so great was the concierge. There was a wonderful young lady there that recommended 2 wonderful resturants for dinners and an amazing dim sum resturant for lunch. We enjoyed San Fran. way [SEP]\n",
+      "[CLS] Was the atmosphere of the tourist areas of san fransisco peaceful? [SEP] TV anyway. What made our stay so great was the concierge. There was a wonderful young lady there that recommended 2 wonderful resturants for dinners and an amazing dim sum resturant for lunch. We enjoyed San Fran. way more then we ever thought we would and we owe that to The Argonaut Hotel and there staff. We can't wait to come visit again. [SEP]\n",
+      "[CLS] How is the employee service on this hotel? [SEP] Spent one night at The Argonaut and wish that we could have stayed longer. We even missed the free wine hour but what the heck, the staff was so pleasant and not in a superficial,'I don't mean it'kind of way. Valet was on top of their game and was very helpful with directions or anything we might need. The roofm was super clean and I was looking for dirt in the areas [SEP]\n",
+      "[CLS] How is the employee service on this hotel? [SEP] a superficial,'I don't mean it'kind of way. Valet was on top of their game and was very helpful with directions or anything we might need. The roofm was super clean and I was looking for dirt in the areas that sometimes don't get the best housekeeping job. I couldn't find as much as a crumb! We were on the fourth floor in a two queen bedded room [SEP]\n",
+      "[CLS] How is the employee service on this hotel? [SEP] was super clean and I was looking for dirt in the areas that sometimes don't get the best housekeeping job. I couldn't find as much as a crumb! We were on the fourth floor in a two queen bedded room as we were traveling with our two children. We were upgraded to a Cannery / Alcatraz view because of our membership in Kimpton's In Touch program but it was really just [SEP]\n",
+      "[CLS] How is the employee service on this hotel? [SEP] were on the fourth floor in a two queen bedded room as we were traveling with our two children. We were upgraded to a Cannery / Alcatraz view because of our membership in Kimpton's In Touch program but it was really just a view of the Mexican restaurant below unless you really, really cranked your head. Still nice to receive. We also had a note welcoming us and a big bottle of water but didn [SEP]\n",
+      "[CLS] How is the employee service on this hotel? [SEP] Kimpton's In Touch program but it was really just a view of the Mexican restaurant below unless you really, really cranked your head. Still nice to receive. We also had a note welcoming us and a big bottle of water but didn't drink it because it didn't say it was complementary. Overall a great experience and I wouldn't hesitate to return whenever I am in the San Francisco area. Great decor [SEP]\n",
+      "[CLS] How is the employee service on this hotel? [SEP] a note welcoming us and a big bottle of water but didn't drink it because it didn't say it was complementary. Overall a great experience and I wouldn't hesitate to return whenever I am in the San Francisco area. Great decor. This was our first stay in a Kimtpon property but I am going to seek them out from now on. [SEP]\n",
+      "[CLS] How is attraction? [SEP] This is a great hotel! I loved the nautical decoration! The room was immaculate, very comfortable and we didn't have any issues with noise. The location is great, especially for a quick stay as you are close to the tourist attractions. There are countless kiosks in the area for different tours and attractions. We took the cable car tour around San Fran and over the Golden Gate Bridge. The ferry to Alcatraz is a 10 - [SEP]\n",
+      "[CLS] How is attraction? [SEP] you are close to the tourist attractions. There are countless kiosks in the area for different tours and attractions. We took the cable car tour around San Fran and over the Golden Gate Bridge. The ferry to Alcatraz is a 10 - 15 minute walk away from the hotel. There are loads of restaurants nearby. The restaurant attached to the hotel was great for breakfast. I would definitely recommend this hotel. [SEP]\n"
+     ]
+    }
+   ],
+   "source": [
+    "context = subjqa[\"train\"][2:6][\"context\"]\n",
+    "question = subjqa[\"train\"][2:6][\"question\"]\n",
+    "\n",
+    "inputs = tokenizer(\n",
+    "    question,\n",
+    "    context,\n",
+    "    max_length=100,\n",
+    "    truncation=\"only_second\",\n",
+    "    stride=50,\n",
+    "    return_overflowing_tokens=True,\n",
+    "    return_offsets_mapping=True,\n",
+    ")\n",
+    "for ids in inputs[\"input_ids\"]:\n",
+    "    print(tokenizer.decode(ids))"
+   ],
+   "metadata": {
+    "collapsed": false,
+    "pycharm": {
+     "name": "#%%\n"
+    }
+   }
+  },
+  {
+   "cell_type": "code",
+   "execution_count": 23,
+   "outputs": [
+    {
+     "name": "stdout",
+     "output_type": "stream",
+     "text": [
+      "The 4 examples gave 18 features.\n",
+      "Here is where each comes from: [0, 1, 1, 1, 1, 1, 1, 1, 1, 1, 2, 2, 2, 2, 2, 2, 3, 3].\n"
+     ]
+    }
+   ],
+   "source": [
+    "print(f\"The 4 examples gave {len(inputs['input_ids'])} features.\")\n",
+    "print(f\"Here is where each comes from: {inputs['overflow_to_sample_mapping']}.\")"
+   ],
+   "metadata": {
+    "collapsed": false,
+    "pycharm": {
+     "name": "#%%\n"
+    }
+   }
+  },
+  {
+   "cell_type": "code",
+   "execution_count": 24,
+   "outputs": [
+    {
+     "name": "stdout",
+     "output_type": "stream",
+     "text": [
+      "[31, 0, 0, 0, 0, 0, 0, 0, 0, 61, 41, 0, 0, 0, 0, 0, 70, 27]\n",
+      "[31, 0, 0, 0, 0, 0, 0, 0, 0, 72, 45, 0, 0, 0, 0, 0, 70, 27]\n"
+     ]
+    }
+   ],
+   "source": [
+    "answers = subjqa[\"train\"][2:6][\"answers\"]\n",
+    "start_positions = []\n",
+    "end_positions = []\n",
+    "\n",
+    "for i, offset in enumerate(inputs[\"offset_mapping\"]):\n",
+    "    sample_idx = inputs[\"overflow_to_sample_mapping\"][i]\n",
+    "    sequence_ids = inputs.sequence_ids(i)\n",
+    "    answer = answers[sample_idx]\n",
+    "\n",
+    "    # Find start and end of context\n",
+    "    context_start = sequence_ids.index(1)\n",
+    "    context_end = len(sequence_ids) - 2\n",
+    "    start_char = 0\n",
+    "    end_char = 0\n",
+    "    if len(answer['answer_start']) > 0:\n",
+    "        for idx, answer_start in enumerate(answer['answer_start']):\n",
+    "            tmp_start_char = answer_start\n",
+    "            tmp_end_char = answer_start + len(answer['text'][idx])\n",
+    "            # Answer inside context\n",
+    "            if offset[context_start][0] <= tmp_start_char and offset[context_end][1] >= tmp_end_char:\n",
+    "                idx_c = context_start\n",
+    "                while idx_c <= context_end and offset[idx_c][0] <= tmp_start_char:\n",
+    "                    idx_c += 1\n",
+    "                start_char = idx_c - 1\n",
+    "\n",
+    "                idx_c = context_end\n",
+    "                while idx_c >= context_start and offset[idx_c][1] >= tmp_end_char:\n",
+    "                    idx_c -= 1\n",
+    "                end_char = idx_c + 1\n",
+    "                break\n",
+    "\n",
+    "    start_positions.append(start_char)\n",
+    "    end_positions.append(end_char)\n",
+    "\n",
+    "print(start_positions)\n",
+    "print(end_positions)"
+   ],
+   "metadata": {
+    "collapsed": false,
+    "pycharm": {
+     "name": "#%%\n"
+    }
+   }
+  },
+  {
+   "cell_type": "code",
+   "execution_count": 25,
+   "outputs": [
+    {
+     "name": "stdout",
+     "output_type": "stream",
+     "text": [
+      "Theoretical answer: 18, labels give: 18\n"
+     ]
+    }
+   ],
+   "source": [
+    "idx = 0\n",
+    "sample_idx = inputs[\"overflow_to_sample_mapping\"][idx]\n",
+    "answer = answers[sample_idx][\"text\"][0]\n",
+    "\n",
+    "start = start_positions[idx]\n",
+    "end = end_positions[idx]\n",
+    "labeled_answer = tokenizer.decode(inputs[\"input_ids\"][idx][start: end + 1])\n",
+    "\n",
+    "print(f\"Theoretical answer: {answer}, labels give: {labeled_answer}\")\n",
+    "del inputs\n"
+   ],
+   "metadata": {
+    "collapsed": false,
+    "pycharm": {
+     "name": "#%%\n"
+    }
+   }
+  },
+  {
+   "cell_type": "markdown",
+   "source": [
+    "Przygotowanie do stride"
+   ],
+   "metadata": {
+    "collapsed": false,
+    "pycharm": {
+     "name": "#%% md\n"
+    }
+   }
+  },
+  {
+   "cell_type": "code",
+   "execution_count": 47,
+   "outputs": [],
+   "source": [
+    "\n",
+    "def preprocess_train_data(examples, max_length=384, stride=128):\n",
+    "    questions = [q.strip() for q in examples[\"question\"]]\n",
+    "    print(tokenizer.__class__.__name__)\n",
+    "    inputs = tokenizer(\n",
+    "        questions,\n",
+    "        examples[\"context\"],\n",
+    "        max_length=max_length,\n",
+    "        truncation=\"only_second\",\n",
+    "        stride=stride,\n",
+    "        return_overflowing_tokens=True,\n",
+    "        return_offsets_mapping=True,\n",
+    "        padding=\"max_length\",\n",
+    "    )\n",
+    "    answers = examples[\"answers\"]\n",
+    "    start_positions = []\n",
+    "    end_positions = []\n",
+    "\n",
+    "    for i, offset in enumerate(inputs[\"offset_mapping\"]):\n",
+    "        sample_idx = inputs[\"overflow_to_sample_mapping\"][i]\n",
+    "        sequence_ids = inputs.sequence_ids(i)\n",
+    "        answer = answers[sample_idx]\n",
+    "\n",
+    "        # Find start and end of context\n",
+    "        context_start = sequence_ids.index(1)\n",
+    "        context_end = len(sequence_ids) - 2\n",
+    "        start_char = 0\n",
+    "        end_char = 0\n",
+    "        if len(answer['answer_start']) > 0:\n",
+    "            for idx, answer_start in enumerate(answer['answer_start']):\n",
+    "                tmp_start_char = answer_start\n",
+    "                tmp_end_char = answer_start + len(answer['text'][idx])\n",
+    "                # Answer inside context\n",
+    "                if offset[context_start][0] <= tmp_start_char and offset[context_end][1] >= tmp_end_char:\n",
+    "                    idx_c = context_start\n",
+    "                    while idx_c <= context_end and offset[idx_c][0] <= tmp_start_char:\n",
+    "                        idx_c += 1\n",
+    "                    start_char = idx_c - 1\n",
+    "\n",
+    "                    idx_c = context_end\n",
+    "                    while idx_c >= context_start and offset[idx_c][1] >= tmp_end_char:\n",
+    "                        idx_c -= 1\n",
+    "                    end_char = idx_c + 1\n",
+    "                    break\n",
+    "\n",
+    "        start_positions.append(start_char)\n",
+    "        end_positions.append(end_char)\n",
+    "    inputs[\"start_positions\"] = start_positions\n",
+    "    inputs[\"end_positions\"] = end_positions\n",
+    "    return inputs"
+   ],
+   "metadata": {
+    "collapsed": false,
+    "pycharm": {
+     "name": "#%%\n"
+    }
+   }
+  },
+  {
+   "cell_type": "code",
+   "execution_count": 27,
+   "outputs": [
+    {
+     "data": {
+      "text/plain": "  0%|          | 0/2 [00:00<?, ?ba/s]",
+      "application/vnd.jupyter.widget-view+json": {
+       "version_major": 2,
+       "version_minor": 0,
+       "model_id": "de09304260654d5aaf6b0801f0264006"
+      },
+      "application/json": {
+       "n": 0,
+       "total": 2,
+       "elapsed": 0.010731697082519531,
+       "ncols": null,
+       "nrows": 12,
+       "prefix": "",
+       "ascii": false,
+       "unit": "ba",
+       "unit_scale": false,
+       "rate": null,
+       "bar_format": null,
+       "postfix": null,
+       "unit_divisor": 1000,
+       "initial": 0,
+       "colour": null
+      }
+     },
+     "metadata": {},
+     "output_type": "display_data"
+    },
+    {
+     "data": {
+      "text/plain": "(1666, 2030)"
+     },
+     "execution_count": 27,
+     "metadata": {},
+     "output_type": "execute_result"
+    }
+   ],
+   "source": [
+    "train_dataset = subjqa[\"train\"].map(\n",
+    "    preprocess_train_data,\n",
+    "    batched=True,\n",
+    "    remove_columns=subjqa[\"train\"].column_names,\n",
+    ")\n",
+    "len(subjqa[\"train\"]), len(train_dataset)"
+   ],
+   "metadata": {
+    "collapsed": false,
+    "pycharm": {
+     "name": "#%%\n"
+    }
+   }
+  },
+  {
+   "cell_type": "code",
+   "execution_count": 48,
+   "outputs": [],
+   "source": [
+    "def preprocess_validation_examples(examples, max_length=384, stride=128):\n",
+    "    questions = [q.strip() for q in examples[\"question\"]]\n",
+    "    print(tokenizer.__class__.__name__)\n",
+    "    inputs = tokenizer(\n",
+    "        questions,\n",
+    "        examples[\"context\"],\n",
+    "        max_length=max_length,\n",
+    "        truncation=\"only_second\",\n",
+    "        stride=stride,\n",
+    "        return_overflowing_tokens=True,\n",
+    "        return_offsets_mapping=True,\n",
+    "        padding=\"max_length\",\n",
+    "    )\n",
+    "\n",
+    "    sample_map = inputs.pop(\"overflow_to_sample_mapping\")\n",
+    "    example_ids = []\n",
+    "\n",
+    "    for i in range(len(inputs[\"input_ids\"])):\n",
+    "        sample_idx = sample_map[i]\n",
+    "        example_ids.append(examples[\"id\"][sample_idx])\n",
+    "\n",
+    "        sequence_ids = inputs.sequence_ids(i)\n",
+    "        offset = inputs[\"offset_mapping\"][i]\n",
+    "        inputs[\"offset_mapping\"][i] = [\n",
+    "            o if sequence_ids[k] == 1 else None for k, o in enumerate(offset)\n",
+    "        ]\n",
+    "\n",
+    "    inputs[\"example_id\"] = example_ids\n",
+    "    return inputs"
+   ],
+   "metadata": {
+    "collapsed": false,
+    "pycharm": {
+     "name": "#%%\n"
+    }
+   }
+  },
+  {
+   "cell_type": "code",
+   "execution_count": 29,
+   "outputs": [
+    {
+     "data": {
+      "text/plain": "  0%|          | 0/1 [00:00<?, ?ba/s]",
+      "application/vnd.jupyter.widget-view+json": {
+       "version_major": 2,
+       "version_minor": 0,
+       "model_id": "a490929957d24cb0baca4341002d3139"
+      },
+      "application/json": {
+       "n": 0,
+       "total": 1,
+       "elapsed": 0.009812355041503906,
+       "ncols": null,
+       "nrows": 12,
+       "prefix": "",
+       "ascii": false,
+       "unit": "ba",
+       "unit_scale": false,
+       "rate": null,
+       "bar_format": null,
+       "postfix": null,
+       "unit_divisor": 1000,
+       "initial": 0,
+       "colour": null
+      }
+     },
+     "metadata": {},
+     "output_type": "display_data"
+    },
+    {
+     "data": {
+      "text/plain": "(265, 327)"
+     },
+     "execution_count": 29,
+     "metadata": {},
+     "output_type": "execute_result"
+    }
+   ],
+   "source": [
+    "validation_dataset = subjqa[\"validation\"].map(\n",
+    "    preprocess_validation_examples,\n",
+    "    batched=True,\n",
+    "    remove_columns=subjqa[\"validation\"].column_names,\n",
+    ")\n",
+    "len(subjqa[\"validation\"]), len(validation_dataset)"
+   ],
+   "metadata": {
+    "collapsed": false,
+    "pycharm": {
+     "name": "#%%\n"
+    }
+   }
+  },
+  {
+   "cell_type": "markdown",
+   "source": [
+    "# Experymenty"
+   ],
+   "metadata": {
+    "collapsed": false,
+    "pycharm": {
+     "name": "#%% md\n"
+    }
+   }
+  },
+  {
+   "cell_type": "code",
+   "execution_count": 30,
+   "outputs": [],
+   "source": [
+    "from tqdm.auto import tqdm\n",
+    "import evaluate\n",
+    "import collections\n",
+    "\n",
+    "device = torch.device(\"cuda\") if torch.cuda.is_available() else torch.device(\"cpu\")\n",
+    "metric = evaluate.load(\"squad\")\n",
+    "n_best = 3\n",
+    "max_answer_length = 50\n",
+    "predicted_answers = []\n",
+    "\n",
+    "\n",
+    "def compute_metrics(start_logits, end_logits, features, examples):\n",
+    "    example_to_features = collections.defaultdict(list)\n",
+    "    for idx, feature in enumerate(features):\n",
+    "        example_to_features[feature[\"example_id\"]].append(idx)\n",
+    "\n",
+    "    predicted_answers = []\n",
+    "    for example in tqdm(examples):\n",
+    "        example_id = example[\"id\"]\n",
+    "        context = example[\"context\"]\n",
+    "        answers = []\n",
+    "\n",
+    "        # Loop through all features associated with that example\n",
+    "        for feature_index in example_to_features[example_id]:\n",
+    "            start_logit = start_logits[feature_index]\n",
+    "            end_logit = end_logits[feature_index]\n",
+    "            offsets = features[feature_index][\"offset_mapping\"]\n",
+    "\n",
+    "            start_indexes = np.argsort(start_logit)[-1: -n_best - 1: -1].tolist()\n",
+    "            end_indexes = np.argsort(end_logit)[-1: -n_best - 1: -1].tolist()\n",
+    "            for start_index in start_indexes:\n",
+    "                for end_index in end_indexes:\n",
+    "                    # Skip answers that are not fully in the context\n",
+    "                    if offsets[start_index] is None or offsets[end_index] is None:\n",
+    "                        continue\n",
+    "                    # Skip answers with a length that is either < 0 or > max_answer_length\n",
+    "                    if (\n",
+    "                            end_index < start_index\n",
+    "                            or end_index - start_index + 1 > max_answer_length\n",
+    "                    ):\n",
+    "                        continue\n",
+    "\n",
+    "                    answer = {\n",
+    "                        \"text\": context[offsets[start_index][0]: offsets[end_index][1]],\n",
+    "                        \"logit_score\": start_logit[start_index] + end_logit[end_index],\n",
+    "                    }\n",
+    "                    answers.append(answer)\n",
+    "\n",
+    "        # Select the answer with the best score\n",
+    "        if len(answers) > 0:\n",
+    "            best_answer = max(answers, key=lambda x: x[\"logit_score\"])\n",
+    "            predicted_answers.append(\n",
+    "                {\"id\": example_id, \"prediction_text\": best_answer[\"text\"]}\n",
+    "            )\n",
+    "        else:\n",
+    "            predicted_answers.append({\"id\": example_id, \"prediction_text\": \"\"})\n",
+    "\n",
+    "    theoretical_answers = [\n",
+    "        {\"id\": ex[\"id\"], \"answers\": {\n",
+    "            'text': ex[\"answers\"]['text']\n",
+    "            if len(ex[\"answers\"]['text']) != 0\n",
+    "            else [\"\"],\n",
+    "            'answer_start': ex[\"answers\"][\"answer_start\"]\n",
+    "            if len(ex[\"answers\"][\"answer_start\"]) != 0\n",
+    "            else [0]\n",
+    "        }} for ex in examples\n",
+    "    ]\n",
+    "    metrics = metric.compute(predictions=predicted_answers, references=theoretical_answers)\n",
+    "    for i in range(3):\n",
+    "        print(\"QUESTION:\\t\", examples[i]['question'])\n",
+    "        print(\"PREDICTED:\", predicted_answers[i]['prediction_text'])\n",
+    "        print(\"ACTUAL:\", theoretical_answers[i]['answers']['text'])\n",
+    "    print(metrics)\n",
+    "\n",
+    "    return predicted_answers, theoretical_answers, metrics"
+   ],
+   "metadata": {
+    "collapsed": false,
+    "pycharm": {
+     "name": "#%%\n"
+    }
+   }
+  },
+  {
+   "cell_type": "code",
+   "execution_count": 50,
+   "outputs": [],
+   "source": [
+    "import numpy as np\n",
+    "\n",
+    "def predict_from_trained():\n",
+    "    eval_set = subjqa[\"validation\"].map(\n",
+    "    preprocess_validation_examples,\n",
+    "    batched=True,\n",
+    "    remove_columns=subjqa[\"validation\"].column_names,\n",
+    "    )\n",
+    "    eval_set_for_model = eval_set.remove_columns([\"example_id\", \"offset_mapping\"])\n",
+    "    eval_set_for_model.set_format(\"torch\")\n",
+    "    batch_size = 8\n",
+    "    all_start_logits = []\n",
+    "    all_end_logits = []\n",
+    "    for i in range(0, int(eval_set_for_model.num_rows / batch_size.__ceil__())+1):\n",
+    "        batch = {k: eval_set_for_model[k][batch_size*i:batch_size*(i+1)].to(device) for k in eval_set_for_model.column_names}\n",
+    "        with torch.no_grad():\n",
+    "            outputs = trained_model(**batch)\n",
+    "        all_start_logits.append(outputs.start_logits.cpu().numpy())\n",
+    "        all_end_logits.append(outputs.end_logits.cpu().numpy())\n",
+    "    start_logits = np.concatenate(all_start_logits, axis=0)\n",
+    "    end_logits = np.concatenate(all_end_logits, axis=0)\n",
+    "    _=compute_metrics(start_logits, end_logits, eval_set, subjqa[\"validation\"])\n"
+   ],
+   "metadata": {
+    "collapsed": false,
+    "pycharm": {
+     "name": "#%%\n"
+    }
+   }
+  },
+  {
+   "cell_type": "markdown",
+   "source": [
+    "### distilbert-base-cased-distilled-squad trained"
+   ],
+   "metadata": {
+    "collapsed": false,
+    "pycharm": {
+     "name": "#%% md\n"
+    }
+   }
+  },
+  {
+   "cell_type": "code",
+   "execution_count": 51,
+   "outputs": [
+    {
+     "name": "stderr",
+     "output_type": "stream",
+     "text": [
+      "Loading cached processed dataset at /home/karo/.cache/huggingface/datasets/subjqa/tripadvisor/1.1.0/e5588f9298ff2d70686a00cc377e4bdccf4e32287459e3c6baf2dc5ab57fe7fd/cache-bfb7935e995d8aee.arrow\n"
+     ]
+    },
+    {
+     "data": {
+      "text/plain": "  0%|          | 0/265 [00:00<?, ?it/s]",
+      "application/vnd.jupyter.widget-view+json": {
+       "version_major": 2,
+       "version_minor": 0,
+       "model_id": "de814472ffd44fdb9167beb5e806c229"
+      },
+      "application/json": {
+       "n": 0,
+       "total": 265,
+       "elapsed": 0.009649038314819336,
+       "ncols": null,
+       "nrows": 12,
+       "prefix": "",
+       "ascii": false,
+       "unit": "it",
+       "unit_scale": false,
+       "rate": null,
+       "bar_format": null,
+       "postfix": null,
+       "unit_divisor": 1000,
+       "initial": 0,
+       "colour": null
+      }
+     },
+     "metadata": {},
+     "output_type": "display_data"
+    },
+    {
+     "name": "stdout",
+     "output_type": "stream",
+     "text": [
+      "QUESTION:\t How was the hotel?\n",
+      "PREDICTED: it is a quality 4-star hotel\n",
+      "ACTUAL: ['excellent hotels']\n",
+      "QUESTION:\t How is the hotel?\n",
+      "PREDICTED: The hotel is clean and neat\n",
+      "ACTUAL: ['The hotel location was great']\n",
+      "QUESTION:\t Is it value for money?\n",
+      "PREDICTED: excellent value for money\n",
+      "ACTUAL: ['And very reasonably priced. Overall, excellent value for money and highly recommended.']\n",
+      "{'exact_match': 10.943396226415095, 'f1': 33.948637559395756}\n"
+     ]
+    }
+   ],
+   "source": [
+    "trained_checkpoint = \"distilbert-base-cased-distilled-squad\"\n",
+    "\n",
+    "tokenizer = AutoTokenizer.from_pretrained(trained_checkpoint)\n",
+    "trained_model = AutoModelForQuestionAnswering.from_pretrained(trained_checkpoint).to(device)\n",
+    "predict_from_trained()\n"
+   ],
+   "metadata": {
+    "collapsed": false,
+    "pycharm": {
+     "name": "#%%\n"
+    }
+   }
+  },
+  {
+   "cell_type": "markdown",
+   "source": [
+    "### Bert Uncased Fine tune"
+   ],
+   "metadata": {
+    "collapsed": false,
+    "pycharm": {
+     "name": "#%% md\n"
+    }
+   }
+  },
+  {
+   "cell_type": "code",
+   "execution_count": 52,
+   "outputs": [
+    {
+     "name": "stderr",
+     "output_type": "stream",
+     "text": [
+      "Some weights of the model checkpoint at bert-base-cased were not used when initializing BertForQuestionAnswering: ['cls.seq_relationship.bias', 'cls.seq_relationship.weight', 'cls.predictions.decoder.weight', 'cls.predictions.transform.LayerNorm.bias', 'cls.predictions.bias', 'cls.predictions.transform.dense.weight', 'cls.predictions.transform.LayerNorm.weight', 'cls.predictions.transform.dense.bias']\n",
+      "- This IS expected if you are initializing BertForQuestionAnswering from the checkpoint of a model trained on another task or with another architecture (e.g. initializing a BertForSequenceClassification model from a BertForPreTraining model).\n",
+      "- This IS NOT expected if you are initializing BertForQuestionAnswering from the checkpoint of a model that you expect to be exactly identical (initializing a BertForSequenceClassification model from a BertForSequenceClassification model).\n",
+      "Some weights of BertForQuestionAnswering were not initialized from the model checkpoint at bert-base-cased and are newly initialized: ['qa_outputs.weight', 'qa_outputs.bias']\n",
+      "You should probably TRAIN this model on a down-stream task to be able to use it for predictions and inference.\n"
+     ]
+    }
+   ],
+   "source": [
+    "model = AutoModelForQuestionAnswering.from_pretrained(model_ckpt)\n",
+    "tokenizer = AutoTokenizer.from_pretrained(model_ckpt)"
+   ],
+   "metadata": {
+    "collapsed": false,
+    "pycharm": {
+     "name": "#%%\n"
+    }
+   }
+  },
+  {
+   "cell_type": "code",
+   "execution_count": 53,
+   "outputs": [
+    {
+     "name": "stderr",
+     "output_type": "stream",
+     "text": [
+      "Using cuda_amp half precision backend\n",
+      "The following columns in the training set don't have a corresponding argument in `BertForQuestionAnswering.forward` and have been ignored: offset_mapping, overflow_to_sample_mapping. If offset_mapping, overflow_to_sample_mapping are not expected by `BertForQuestionAnswering.forward`,  you can safely ignore this message.\n",
+      "/home/karo/nlp-project/venv/lib/python3.10/site-packages/transformers/optimization.py:306: FutureWarning: This implementation of AdamW is deprecated and will be removed in a future version. Use the PyTorch implementation torch.optim.AdamW instead, or set `no_deprecation_warning=True` to disable this warning\n",
+      "  warnings.warn(\n",
+      "***** Running training *****\n",
+      "  Num examples = 2030\n",
+      "  Num Epochs = 3\n",
+      "  Instantaneous batch size per device = 8\n",
+      "  Total train batch size (w. parallel, distributed & accumulation) = 8\n",
+      "  Gradient Accumulation steps = 1\n",
+      "  Total optimization steps = 762\n"
+     ]
+    },
+    {
+     "data": {
+      "text/plain": "<IPython.core.display.HTML object>",
+      "text/html": "\n    <div>\n      \n      <progress value='2' max='762' style='width:300px; height:20px; vertical-align: middle;'></progress>\n      [  2/762 : < :, Epoch 0.00/3]\n    </div>\n    <table border=\"1\" class=\"dataframe\">\n  <thead>\n <tr style=\"text-align: left;\">\n      <th>Step</th>\n      <th>Training Loss</th>\n    </tr>\n  </thead>\n  <tbody>\n  </tbody>\n</table><p>"
+     },
+     "metadata": {},
+     "output_type": "display_data"
+    },
+    {
+     "name": "stderr",
+     "output_type": "stream",
+     "text": [
+      "Saving model checkpoint to data/bert-finetuned-subjqa/checkpoint-254\n",
+      "Configuration saved in data/bert-finetuned-subjqa/checkpoint-254/config.json\n",
+      "Model weights saved in data/bert-finetuned-subjqa/checkpoint-254/pytorch_model.bin\n",
+      "tokenizer config file saved in data/bert-finetuned-subjqa/checkpoint-254/tokenizer_config.json\n",
+      "Special tokens file saved in data/bert-finetuned-subjqa/checkpoint-254/special_tokens_map.json\n",
+      "Saving model checkpoint to data/bert-finetuned-subjqa/checkpoint-508\n",
+      "Configuration saved in data/bert-finetuned-subjqa/checkpoint-508/config.json\n",
+      "Model weights saved in data/bert-finetuned-subjqa/checkpoint-508/pytorch_model.bin\n",
+      "tokenizer config file saved in data/bert-finetuned-subjqa/checkpoint-508/tokenizer_config.json\n",
+      "Special tokens file saved in data/bert-finetuned-subjqa/checkpoint-508/special_tokens_map.json\n",
+      "Saving model checkpoint to data/bert-finetuned-subjqa/checkpoint-762\n",
+      "Configuration saved in data/bert-finetuned-subjqa/checkpoint-762/config.json\n",
+      "Model weights saved in data/bert-finetuned-subjqa/checkpoint-762/pytorch_model.bin\n",
+      "tokenizer config file saved in data/bert-finetuned-subjqa/checkpoint-762/tokenizer_config.json\n",
+      "Special tokens file saved in data/bert-finetuned-subjqa/checkpoint-762/special_tokens_map.json\n",
+      "\n",
+      "\n",
+      "Training completed. Do not forget to share your model on huggingface.co/models =)\n",
+      "\n",
+      "\n"
+     ]
+    },
+    {
+     "data": {
+      "text/plain": "TrainOutput(global_step=762, training_loss=0.7420042443463183, metrics={'train_runtime': 336.3364, 'train_samples_per_second': 18.107, 'train_steps_per_second': 2.266, 'total_flos': 1193472936391680.0, 'train_loss': 0.7420042443463183, 'epoch': 3.0})"
+     },
+     "execution_count": 53,
+     "metadata": {},
+     "output_type": "execute_result"
+    }
+   ],
+   "source": [
+    "from transformers import TrainingArguments\n",
+    "from transformers import Trainer\n",
+    "\n",
+    "\n",
+    "args = TrainingArguments(\n",
+    "    output_dir=\"data/bert-finetuned-subjqa\",\n",
+    "    overwrite_output_dir=True,\n",
+    "    evaluation_strategy=\"no\",\n",
+    "    save_strategy=\"epoch\",\n",
+    "    learning_rate=2e-5,\n",
+    "    num_train_epochs=3,\n",
+    "    weight_decay=0.01,\n",
+    "    fp16=True,\n",
+    ")\n",
+    "trainer = Trainer(\n",
+    "    model=model,\n",
+    "    args=args,\n",
+    "    train_dataset=train_dataset,\n",
+    "    eval_dataset=validation_dataset,\n",
+    "    tokenizer=tokenizer,\n",
+    ")\n",
+    "trainer.train()"
+   ],
+   "metadata": {
+    "collapsed": false,
+    "pycharm": {
+     "name": "#%%\n"
+    }
+   }
+  },
+  {
+   "cell_type": "code",
+   "execution_count": 54,
+   "outputs": [
+    {
+     "name": "stderr",
+     "output_type": "stream",
+     "text": [
+      "The following columns in the test set don't have a corresponding argument in `BertForQuestionAnswering.forward` and have been ignored: example_id, offset_mapping. If example_id, offset_mapping are not expected by `BertForQuestionAnswering.forward`,  you can safely ignore this message.\n",
+      "***** Running Prediction *****\n",
+      "  Num examples = 327\n",
+      "  Batch size = 8\n"
+     ]
+    },
+    {
+     "data": {
+      "text/plain": "<IPython.core.display.HTML object>",
+      "text/html": "\n    <div>\n      \n      <progress value='1' max='41' style='width:300px; height:20px; vertical-align: middle;'></progress>\n      [ 1/41 : < :]\n    </div>\n    "
+     },
+     "metadata": {},
+     "output_type": "display_data"
+    },
+    {
+     "data": {
+      "text/plain": "  0%|          | 0/265 [00:00<?, ?it/s]",
+      "application/vnd.jupyter.widget-view+json": {
+       "version_major": 2,
+       "version_minor": 0,
+       "model_id": "99e405e6118f4a57bdd65b16b75fbc55"
+      },
+      "application/json": {
+       "n": 0,
+       "total": 265,
+       "elapsed": 0.010312318801879883,
+       "ncols": null,
+       "nrows": 12,
+       "prefix": "",
+       "ascii": false,
+       "unit": "it",
+       "unit_scale": false,
+       "rate": null,
+       "bar_format": null,
+       "postfix": null,
+       "unit_divisor": 1000,
+       "initial": 0,
+       "colour": null
+      }
+     },
+     "metadata": {},
+     "output_type": "display_data"
+    },
+    {
+     "name": "stdout",
+     "output_type": "stream",
+     "text": [
+      "QUESTION:\t How was the hotel?\n",
+      "PREDICTED: Great setting at the end of the wharf\n",
+      "ACTUAL: ['excellent hotels']\n",
+      "QUESTION:\t How is the hotel?\n",
+      "PREDICTED: My wife and I took two trips to San Fran in 2004 and stayed at the best western both times\n",
+      "ACTUAL: ['The hotel location was great']\n",
+      "QUESTION:\t Is it value for money?\n",
+      "PREDICTED: excellent value for money\n",
+      "ACTUAL: ['And very reasonably priced. Overall, excellent value for money and highly recommended.']\n",
+      "{'exact_match': 13.584905660377359, 'f1': 30.577012885313252}\n"
+     ]
+    }
+   ],
+   "source": [
+    "predictions, _, _ = trainer.predict(validation_dataset)\n",
+    "start_logits, end_logits = predictions\n",
+    "_=compute_metrics(start_logits, end_logits, validation_dataset, subjqa[\"validation\"])"
+   ],
+   "metadata": {
+    "collapsed": false,
+    "pycharm": {
+     "name": "#%%\n"
+    }
+   }
+  },
+  {
+   "cell_type": "code",
+   "execution_count": 55,
+   "outputs": [
+    {
+     "name": "stderr",
+     "output_type": "stream",
+     "text": [
+      "loading configuration file https://huggingface.co/deepset/roberta-base-squad2/resolve/main/config.json from cache at /home/karo/.cache/huggingface/transformers/c40d0abb589629c48763f271020d0b1f602f5208c432c0874d420491ed37e28b.122ed338b3591c07dba452777c59ff52330edb340d3d56d67aa9117ad9905673\n",
+      "Model config RobertaConfig {\n",
+      "  \"_name_or_path\": \"deepset/roberta-base-squad2\",\n",
+      "  \"architectures\": [\n",
+      "    \"RobertaForQuestionAnswering\"\n",
+      "  ],\n",
+      "  \"attention_probs_dropout_prob\": 0.1,\n",
+      "  \"bos_token_id\": 0,\n",
+      "  \"classifier_dropout\": null,\n",
+      "  \"eos_token_id\": 2,\n",
+      "  \"gradient_checkpointing\": false,\n",
+      "  \"hidden_act\": \"gelu\",\n",
+      "  \"hidden_dropout_prob\": 0.1,\n",
+      "  \"hidden_size\": 768,\n",
+      "  \"initializer_range\": 0.02,\n",
+      "  \"intermediate_size\": 3072,\n",
+      "  \"language\": \"english\",\n",
+      "  \"layer_norm_eps\": 1e-05,\n",
+      "  \"max_position_embeddings\": 514,\n",
+      "  \"model_type\": \"roberta\",\n",
+      "  \"name\": \"Roberta\",\n",
+      "  \"num_attention_heads\": 12,\n",
+      "  \"num_hidden_layers\": 12,\n",
+      "  \"pad_token_id\": 1,\n",
+      "  \"position_embedding_type\": \"absolute\",\n",
+      "  \"transformers_version\": \"4.21.1\",\n",
+      "  \"type_vocab_size\": 1,\n",
+      "  \"use_cache\": true,\n",
+      "  \"vocab_size\": 50265\n",
+      "}\n",
+      "\n",
+      "loading file https://huggingface.co/deepset/roberta-base-squad2/resolve/main/vocab.json from cache at /home/karo/.cache/huggingface/transformers/81c80edb4c6cefa5cae64ccfdb34b3b309ecaf60da99da7cd1c17e24a5d36eb5.647b4548b6d9ea817e82e7a9231a320231a1c9ea24053cc9e758f3fe68216f05\n",
+      "loading file https://huggingface.co/deepset/roberta-base-squad2/resolve/main/merges.txt from cache at /home/karo/.cache/huggingface/transformers/b87d46371731376b11768b7839b1a5938a4f77d6bd2d9b683f167df0026af432.5d12962c5ee615a4c803841266e9c3be9a691a924f72d395d3a6c6c81157788b\n",
+      "loading file https://huggingface.co/deepset/roberta-base-squad2/resolve/main/tokenizer.json from cache at None\n",
+      "loading file https://huggingface.co/deepset/roberta-base-squad2/resolve/main/added_tokens.json from cache at None\n",
+      "loading file https://huggingface.co/deepset/roberta-base-squad2/resolve/main/special_tokens_map.json from cache at /home/karo/.cache/huggingface/transformers/c9d2c178fac8d40234baa1833a3b1903d393729bf93ea34da247c07db24900d0.cb2244924ab24d706b02fd7fcedaea4531566537687a539ebb94db511fd122a0\n",
+      "loading file https://huggingface.co/deepset/roberta-base-squad2/resolve/main/tokenizer_config.json from cache at /home/karo/.cache/huggingface/transformers/e8a600814b69e3ee74bb4a7398cc6fef9812475010f16a6c9f151b2c2772b089.451739a2f3b82c3375da0dfc6af295bedc4567373b171f514dd09a4cc4b31513\n",
+      "loading configuration file https://huggingface.co/deepset/roberta-base-squad2/resolve/main/config.json from cache at /home/karo/.cache/huggingface/transformers/c40d0abb589629c48763f271020d0b1f602f5208c432c0874d420491ed37e28b.122ed338b3591c07dba452777c59ff52330edb340d3d56d67aa9117ad9905673\n",
+      "Model config RobertaConfig {\n",
+      "  \"_name_or_path\": \"deepset/roberta-base-squad2\",\n",
+      "  \"architectures\": [\n",
+      "    \"RobertaForQuestionAnswering\"\n",
+      "  ],\n",
+      "  \"attention_probs_dropout_prob\": 0.1,\n",
+      "  \"bos_token_id\": 0,\n",
+      "  \"classifier_dropout\": null,\n",
+      "  \"eos_token_id\": 2,\n",
+      "  \"gradient_checkpointing\": false,\n",
+      "  \"hidden_act\": \"gelu\",\n",
+      "  \"hidden_dropout_prob\": 0.1,\n",
+      "  \"hidden_size\": 768,\n",
+      "  \"initializer_range\": 0.02,\n",
+      "  \"intermediate_size\": 3072,\n",
+      "  \"language\": \"english\",\n",
+      "  \"layer_norm_eps\": 1e-05,\n",
+      "  \"max_position_embeddings\": 514,\n",
+      "  \"model_type\": \"roberta\",\n",
+      "  \"name\": \"Roberta\",\n",
+      "  \"num_attention_heads\": 12,\n",
+      "  \"num_hidden_layers\": 12,\n",
+      "  \"pad_token_id\": 1,\n",
+      "  \"position_embedding_type\": \"absolute\",\n",
+      "  \"transformers_version\": \"4.21.1\",\n",
+      "  \"type_vocab_size\": 1,\n",
+      "  \"use_cache\": true,\n",
+      "  \"vocab_size\": 50265\n",
+      "}\n",
+      "\n",
+      "loading configuration file https://huggingface.co/deepset/roberta-base-squad2/resolve/main/config.json from cache at /home/karo/.cache/huggingface/transformers/c40d0abb589629c48763f271020d0b1f602f5208c432c0874d420491ed37e28b.122ed338b3591c07dba452777c59ff52330edb340d3d56d67aa9117ad9905673\n",
+      "Model config RobertaConfig {\n",
+      "  \"_name_or_path\": \"deepset/roberta-base-squad2\",\n",
+      "  \"architectures\": [\n",
+      "    \"RobertaForQuestionAnswering\"\n",
+      "  ],\n",
+      "  \"attention_probs_dropout_prob\": 0.1,\n",
+      "  \"bos_token_id\": 0,\n",
+      "  \"classifier_dropout\": null,\n",
+      "  \"eos_token_id\": 2,\n",
+      "  \"gradient_checkpointing\": false,\n",
+      "  \"hidden_act\": \"gelu\",\n",
+      "  \"hidden_dropout_prob\": 0.1,\n",
+      "  \"hidden_size\": 768,\n",
+      "  \"initializer_range\": 0.02,\n",
+      "  \"intermediate_size\": 3072,\n",
+      "  \"language\": \"english\",\n",
+      "  \"layer_norm_eps\": 1e-05,\n",
+      "  \"max_position_embeddings\": 514,\n",
+      "  \"model_type\": \"roberta\",\n",
+      "  \"name\": \"Roberta\",\n",
+      "  \"num_attention_heads\": 12,\n",
+      "  \"num_hidden_layers\": 12,\n",
+      "  \"pad_token_id\": 1,\n",
+      "  \"position_embedding_type\": \"absolute\",\n",
+      "  \"transformers_version\": \"4.21.1\",\n",
+      "  \"type_vocab_size\": 1,\n",
+      "  \"use_cache\": true,\n",
+      "  \"vocab_size\": 50265\n",
+      "}\n",
+      "\n",
+      "loading configuration file https://huggingface.co/deepset/roberta-base-squad2/resolve/main/config.json from cache at /home/karo/.cache/huggingface/transformers/c40d0abb589629c48763f271020d0b1f602f5208c432c0874d420491ed37e28b.122ed338b3591c07dba452777c59ff52330edb340d3d56d67aa9117ad9905673\n",
+      "Model config RobertaConfig {\n",
+      "  \"_name_or_path\": \"deepset/roberta-base-squad2\",\n",
+      "  \"architectures\": [\n",
+      "    \"RobertaForQuestionAnswering\"\n",
+      "  ],\n",
+      "  \"attention_probs_dropout_prob\": 0.1,\n",
+      "  \"bos_token_id\": 0,\n",
+      "  \"classifier_dropout\": null,\n",
+      "  \"eos_token_id\": 2,\n",
+      "  \"gradient_checkpointing\": false,\n",
+      "  \"hidden_act\": \"gelu\",\n",
+      "  \"hidden_dropout_prob\": 0.1,\n",
+      "  \"hidden_size\": 768,\n",
+      "  \"initializer_range\": 0.02,\n",
+      "  \"intermediate_size\": 3072,\n",
+      "  \"language\": \"english\",\n",
+      "  \"layer_norm_eps\": 1e-05,\n",
+      "  \"max_position_embeddings\": 514,\n",
+      "  \"model_type\": \"roberta\",\n",
+      "  \"name\": \"Roberta\",\n",
+      "  \"num_attention_heads\": 12,\n",
+      "  \"num_hidden_layers\": 12,\n",
+      "  \"pad_token_id\": 1,\n",
+      "  \"position_embedding_type\": \"absolute\",\n",
+      "  \"transformers_version\": \"4.21.1\",\n",
+      "  \"type_vocab_size\": 1,\n",
+      "  \"use_cache\": true,\n",
+      "  \"vocab_size\": 50265\n",
+      "}\n",
+      "\n",
+      "loading weights file https://huggingface.co/deepset/roberta-base-squad2/resolve/main/pytorch_model.bin from cache at /home/karo/.cache/huggingface/transformers/eac3273a8097dda671e3bea1db32c616e74f36a306c65b4858171c98d6db83e9.084aa7284f3a51fa1c8f0641aa04c47d366fbd18711f29d0a995693cfdbc9c9e\n",
+      "All model checkpoint weights were used when initializing RobertaForQuestionAnswering.\n",
+      "\n",
+      "All the weights of RobertaForQuestionAnswering were initialized from the model checkpoint at deepset/roberta-base-squad2.\n",
+      "If your task is similar to the task the model of the checkpoint was trained on, you can already use RobertaForQuestionAnswering for predictions without further training.\n"
+     ]
+    },
+    {
+     "data": {
+      "text/plain": "  0%|          | 0/1 [00:00<?, ?ba/s]",
+      "application/vnd.jupyter.widget-view+json": {
+       "version_major": 2,
+       "version_minor": 0,
+       "model_id": "5e883ae315504badb1cd5c1bdf3a957d"
+      },
+      "application/json": {
+       "n": 0,
+       "total": 1,
+       "elapsed": 0.009602069854736328,
+       "ncols": null,
+       "nrows": 12,
+       "prefix": "",
+       "ascii": false,
+       "unit": "ba",
+       "unit_scale": false,
+       "rate": null,
+       "bar_format": null,
+       "postfix": null,
+       "unit_divisor": 1000,
+       "initial": 0,
+       "colour": null
+      }
+     },
+     "metadata": {},
+     "output_type": "display_data"
+    },
+    {
+     "name": "stdout",
+     "output_type": "stream",
+     "text": [
+      "RobertaTokenizerFast\n"
+     ]
+    },
+    {
+     "data": {
+      "text/plain": "  0%|          | 0/265 [00:00<?, ?it/s]",
+      "application/vnd.jupyter.widget-view+json": {
+       "version_major": 2,
+       "version_minor": 0,
+       "model_id": "0b25cfa4cc1243459d19da7f8f3a0ed6"
+      },
+      "application/json": {
+       "n": 0,
+       "total": 265,
+       "elapsed": 0.00951838493347168,
+       "ncols": null,
+       "nrows": 12,
+       "prefix": "",
+       "ascii": false,
+       "unit": "it",
+       "unit_scale": false,
+       "rate": null,
+       "bar_format": null,
+       "postfix": null,
+       "unit_divisor": 1000,
+       "initial": 0,
+       "colour": null
+      }
+     },
+     "metadata": {},
+     "output_type": "display_data"
+    },
+    {
+     "name": "stdout",
+     "output_type": "stream",
+     "text": [
+      "QUESTION:\t How was the hotel?\n",
+      "PREDICTED: No complaints - really happy with what this hotel offered for the price\n",
+      "ACTUAL: ['excellent hotels']\n",
+      "QUESTION:\t How is the hotel?\n",
+      "PREDICTED: clean and neat\n",
+      "ACTUAL: ['The hotel location was great']\n",
+      "QUESTION:\t Is it value for money?\n",
+      "PREDICTED: excellent value for money\n",
+      "ACTUAL: ['And very reasonably priced. Overall, excellent value for money and highly recommended.']\n",
+      "{'exact_match': 11.69811320754717, 'f1': 37.040671268633204}\n"
+     ]
+    }
+   ],
+   "source": [
+    "trained_checkpoint = \"deepset/roberta-base-squad2\"\n",
+    "\n",
+    "tokenizer = AutoTokenizer.from_pretrained(trained_checkpoint)\n",
+    "trained_model = AutoModelForQuestionAnswering.from_pretrained(trained_checkpoint).to(device)\n",
+    "predict_from_trained()"
+   ],
+   "metadata": {
+    "collapsed": false,
+    "pycharm": {
+     "name": "#%%\n"
+    }
+   }
+  },
+  {
+   "cell_type": "code",
+   "execution_count": 56,
+   "outputs": [
+    {
+     "name": "stderr",
+     "output_type": "stream",
+     "text": [
+      "loading configuration file https://huggingface.co/deepset/roberta-base-squad2/resolve/main/config.json from cache at /home/karo/.cache/huggingface/transformers/c40d0abb589629c48763f271020d0b1f602f5208c432c0874d420491ed37e28b.122ed338b3591c07dba452777c59ff52330edb340d3d56d67aa9117ad9905673\n",
+      "Model config RobertaConfig {\n",
+      "  \"_name_or_path\": \"deepset/roberta-base-squad2\",\n",
+      "  \"architectures\": [\n",
+      "    \"RobertaForQuestionAnswering\"\n",
+      "  ],\n",
+      "  \"attention_probs_dropout_prob\": 0.1,\n",
+      "  \"bos_token_id\": 0,\n",
+      "  \"classifier_dropout\": null,\n",
+      "  \"eos_token_id\": 2,\n",
+      "  \"gradient_checkpointing\": false,\n",
+      "  \"hidden_act\": \"gelu\",\n",
+      "  \"hidden_dropout_prob\": 0.1,\n",
+      "  \"hidden_size\": 768,\n",
+      "  \"initializer_range\": 0.02,\n",
+      "  \"intermediate_size\": 3072,\n",
+      "  \"language\": \"english\",\n",
+      "  \"layer_norm_eps\": 1e-05,\n",
+      "  \"max_position_embeddings\": 514,\n",
+      "  \"model_type\": \"roberta\",\n",
+      "  \"name\": \"Roberta\",\n",
+      "  \"num_attention_heads\": 12,\n",
+      "  \"num_hidden_layers\": 12,\n",
+      "  \"pad_token_id\": 1,\n",
+      "  \"position_embedding_type\": \"absolute\",\n",
+      "  \"transformers_version\": \"4.21.1\",\n",
+      "  \"type_vocab_size\": 1,\n",
+      "  \"use_cache\": true,\n",
+      "  \"vocab_size\": 50265\n",
+      "}\n",
+      "\n",
+      "loading weights file https://huggingface.co/deepset/roberta-base-squad2/resolve/main/pytorch_model.bin from cache at /home/karo/.cache/huggingface/transformers/eac3273a8097dda671e3bea1db32c616e74f36a306c65b4858171c98d6db83e9.084aa7284f3a51fa1c8f0641aa04c47d366fbd18711f29d0a995693cfdbc9c9e\n",
+      "All model checkpoint weights were used when initializing RobertaForQuestionAnswering.\n",
+      "\n",
+      "All the weights of RobertaForQuestionAnswering were initialized from the model checkpoint at deepset/roberta-base-squad2.\n",
+      "If your task is similar to the task the model of the checkpoint was trained on, you can already use RobertaForQuestionAnswering for predictions without further training.\n",
+      "loading configuration file https://huggingface.co/deepset/roberta-base-squad2/resolve/main/config.json from cache at /home/karo/.cache/huggingface/transformers/c40d0abb589629c48763f271020d0b1f602f5208c432c0874d420491ed37e28b.122ed338b3591c07dba452777c59ff52330edb340d3d56d67aa9117ad9905673\n",
+      "Model config RobertaConfig {\n",
+      "  \"_name_or_path\": \"deepset/roberta-base-squad2\",\n",
+      "  \"architectures\": [\n",
+      "    \"RobertaForQuestionAnswering\"\n",
+      "  ],\n",
+      "  \"attention_probs_dropout_prob\": 0.1,\n",
+      "  \"bos_token_id\": 0,\n",
+      "  \"classifier_dropout\": null,\n",
+      "  \"eos_token_id\": 2,\n",
+      "  \"gradient_checkpointing\": false,\n",
+      "  \"hidden_act\": \"gelu\",\n",
+      "  \"hidden_dropout_prob\": 0.1,\n",
+      "  \"hidden_size\": 768,\n",
+      "  \"initializer_range\": 0.02,\n",
+      "  \"intermediate_size\": 3072,\n",
+      "  \"language\": \"english\",\n",
+      "  \"layer_norm_eps\": 1e-05,\n",
+      "  \"max_position_embeddings\": 514,\n",
+      "  \"model_type\": \"roberta\",\n",
+      "  \"name\": \"Roberta\",\n",
+      "  \"num_attention_heads\": 12,\n",
+      "  \"num_hidden_layers\": 12,\n",
+      "  \"pad_token_id\": 1,\n",
+      "  \"position_embedding_type\": \"absolute\",\n",
+      "  \"transformers_version\": \"4.21.1\",\n",
+      "  \"type_vocab_size\": 1,\n",
+      "  \"use_cache\": true,\n",
+      "  \"vocab_size\": 50265\n",
+      "}\n",
+      "\n",
+      "loading file https://huggingface.co/deepset/roberta-base-squad2/resolve/main/vocab.json from cache at /home/karo/.cache/huggingface/transformers/81c80edb4c6cefa5cae64ccfdb34b3b309ecaf60da99da7cd1c17e24a5d36eb5.647b4548b6d9ea817e82e7a9231a320231a1c9ea24053cc9e758f3fe68216f05\n",
+      "loading file https://huggingface.co/deepset/roberta-base-squad2/resolve/main/merges.txt from cache at /home/karo/.cache/huggingface/transformers/b87d46371731376b11768b7839b1a5938a4f77d6bd2d9b683f167df0026af432.5d12962c5ee615a4c803841266e9c3be9a691a924f72d395d3a6c6c81157788b\n",
+      "loading file https://huggingface.co/deepset/roberta-base-squad2/resolve/main/tokenizer.json from cache at None\n",
+      "loading file https://huggingface.co/deepset/roberta-base-squad2/resolve/main/added_tokens.json from cache at None\n",
+      "loading file https://huggingface.co/deepset/roberta-base-squad2/resolve/main/special_tokens_map.json from cache at /home/karo/.cache/huggingface/transformers/c9d2c178fac8d40234baa1833a3b1903d393729bf93ea34da247c07db24900d0.cb2244924ab24d706b02fd7fcedaea4531566537687a539ebb94db511fd122a0\n",
+      "loading file https://huggingface.co/deepset/roberta-base-squad2/resolve/main/tokenizer_config.json from cache at /home/karo/.cache/huggingface/transformers/e8a600814b69e3ee74bb4a7398cc6fef9812475010f16a6c9f151b2c2772b089.451739a2f3b82c3375da0dfc6af295bedc4567373b171f514dd09a4cc4b31513\n",
+      "loading configuration file https://huggingface.co/deepset/roberta-base-squad2/resolve/main/config.json from cache at /home/karo/.cache/huggingface/transformers/c40d0abb589629c48763f271020d0b1f602f5208c432c0874d420491ed37e28b.122ed338b3591c07dba452777c59ff52330edb340d3d56d67aa9117ad9905673\n",
+      "Model config RobertaConfig {\n",
+      "  \"_name_or_path\": \"deepset/roberta-base-squad2\",\n",
+      "  \"architectures\": [\n",
+      "    \"RobertaForQuestionAnswering\"\n",
+      "  ],\n",
+      "  \"attention_probs_dropout_prob\": 0.1,\n",
+      "  \"bos_token_id\": 0,\n",
+      "  \"classifier_dropout\": null,\n",
+      "  \"eos_token_id\": 2,\n",
+      "  \"gradient_checkpointing\": false,\n",
+      "  \"hidden_act\": \"gelu\",\n",
+      "  \"hidden_dropout_prob\": 0.1,\n",
+      "  \"hidden_size\": 768,\n",
+      "  \"initializer_range\": 0.02,\n",
+      "  \"intermediate_size\": 3072,\n",
+      "  \"language\": \"english\",\n",
+      "  \"layer_norm_eps\": 1e-05,\n",
+      "  \"max_position_embeddings\": 514,\n",
+      "  \"model_type\": \"roberta\",\n",
+      "  \"name\": \"Roberta\",\n",
+      "  \"num_attention_heads\": 12,\n",
+      "  \"num_hidden_layers\": 12,\n",
+      "  \"pad_token_id\": 1,\n",
+      "  \"position_embedding_type\": \"absolute\",\n",
+      "  \"transformers_version\": \"4.21.1\",\n",
+      "  \"type_vocab_size\": 1,\n",
+      "  \"use_cache\": true,\n",
+      "  \"vocab_size\": 50265\n",
+      "}\n",
+      "\n",
+      "loading configuration file https://huggingface.co/deepset/roberta-base-squad2/resolve/main/config.json from cache at /home/karo/.cache/huggingface/transformers/c40d0abb589629c48763f271020d0b1f602f5208c432c0874d420491ed37e28b.122ed338b3591c07dba452777c59ff52330edb340d3d56d67aa9117ad9905673\n",
+      "Model config RobertaConfig {\n",
+      "  \"_name_or_path\": \"deepset/roberta-base-squad2\",\n",
+      "  \"architectures\": [\n",
+      "    \"RobertaForQuestionAnswering\"\n",
+      "  ],\n",
+      "  \"attention_probs_dropout_prob\": 0.1,\n",
+      "  \"bos_token_id\": 0,\n",
+      "  \"classifier_dropout\": null,\n",
+      "  \"eos_token_id\": 2,\n",
+      "  \"gradient_checkpointing\": false,\n",
+      "  \"hidden_act\": \"gelu\",\n",
+      "  \"hidden_dropout_prob\": 0.1,\n",
+      "  \"hidden_size\": 768,\n",
+      "  \"initializer_range\": 0.02,\n",
+      "  \"intermediate_size\": 3072,\n",
+      "  \"language\": \"english\",\n",
+      "  \"layer_norm_eps\": 1e-05,\n",
+      "  \"max_position_embeddings\": 514,\n",
+      "  \"model_type\": \"roberta\",\n",
+      "  \"name\": \"Roberta\",\n",
+      "  \"num_attention_heads\": 12,\n",
+      "  \"num_hidden_layers\": 12,\n",
+      "  \"pad_token_id\": 1,\n",
+      "  \"position_embedding_type\": \"absolute\",\n",
+      "  \"transformers_version\": \"4.21.1\",\n",
+      "  \"type_vocab_size\": 1,\n",
+      "  \"use_cache\": true,\n",
+      "  \"vocab_size\": 50265\n",
+      "}\n",
+      "\n"
+     ]
+    },
+    {
+     "data": {
+      "text/plain": "  0%|          | 0/2 [00:00<?, ?ba/s]",
+      "application/vnd.jupyter.widget-view+json": {
+       "version_major": 2,
+       "version_minor": 0,
+       "model_id": "562ecfd319d34d7b904fb8ace5b715d8"
+      },
+      "application/json": {
+       "n": 0,
+       "total": 2,
+       "elapsed": 0.009464502334594727,
+       "ncols": null,
+       "nrows": 12,
+       "prefix": "",
+       "ascii": false,
+       "unit": "ba",
+       "unit_scale": false,
+       "rate": null,
+       "bar_format": null,
+       "postfix": null,
+       "unit_divisor": 1000,
+       "initial": 0,
+       "colour": null
+      }
+     },
+     "metadata": {},
+     "output_type": "display_data"
+    },
+    {
+     "name": "stdout",
+     "output_type": "stream",
+     "text": [
+      "RobertaTokenizerFast\n",
+      "RobertaTokenizerFast\n",
+      "1666 1995\n"
+     ]
+    },
+    {
+     "data": {
+      "text/plain": "  0%|          | 0/1 [00:00<?, ?ba/s]",
+      "application/vnd.jupyter.widget-view+json": {
+       "version_major": 2,
+       "version_minor": 0,
+       "model_id": "0a5a6b6a04854ee2ab05b0b01c71db44"
+      },
+      "application/json": {
+       "n": 0,
+       "total": 1,
+       "elapsed": 0.009068012237548828,
+       "ncols": null,
+       "nrows": 12,
+       "prefix": "",
+       "ascii": false,
+       "unit": "ba",
+       "unit_scale": false,
+       "rate": null,
+       "bar_format": null,
+       "postfix": null,
+       "unit_divisor": 1000,
+       "initial": 0,
+       "colour": null
+      }
+     },
+     "metadata": {},
+     "output_type": "display_data"
+    },
+    {
+     "name": "stdout",
+     "output_type": "stream",
+     "text": [
+      "RobertaTokenizerFast\n",
+      "265 321\n"
+     ]
+    }
+   ],
+   "source": [
+    "model_ckpt = \"deepset/roberta-base-squad2\"\n",
+    "\n",
+    "model = AutoModelForQuestionAnswering.from_pretrained(model_ckpt)\n",
+    "tokenizer = AutoTokenizer.from_pretrained(model_ckpt)\n",
+    "\n",
+    "roberta_train_dataset = subjqa[\"train\"].map(\n",
+    "    preprocess_train_data,\n",
+    "    batched=True,\n",
+    "    remove_columns=subjqa[\"train\"].column_names,\n",
+    ")\n",
+    "\n",
+    "print(len(subjqa[\"train\"]), len(roberta_train_dataset))\n",
+    "roberta_validation_dataset = subjqa[\"validation\"].map(\n",
+    "    preprocess_validation_examples,\n",
+    "    batched=True,\n",
+    "    remove_columns=subjqa[\"validation\"].column_names,\n",
+    ")\n",
+    "print(len(subjqa[\"validation\"]), len(roberta_validation_dataset))\n"
+   ],
+   "metadata": {
+    "collapsed": false,
+    "pycharm": {
+     "name": "#%%\n"
+    }
+   }
+  },
+  {
+   "cell_type": "code",
+   "execution_count": 57,
+   "outputs": [
+    {
+     "name": "stderr",
+     "output_type": "stream",
+     "text": [
+      "PyTorch: setting up devices\n",
+      "The default value for the training argument `--report_to` will change in v5 (from all installed integrations to none). In v5, you will need to use `--report_to all` to get the same behavior as now. You should start updating your code and make this info disappear :-).\n",
+      "Using cuda_amp half precision backend\n",
+      "The following columns in the training set don't have a corresponding argument in `RobertaForQuestionAnswering.forward` and have been ignored: offset_mapping, overflow_to_sample_mapping. If offset_mapping, overflow_to_sample_mapping are not expected by `RobertaForQuestionAnswering.forward`,  you can safely ignore this message.\n",
+      "/home/karo/nlp-project/venv/lib/python3.10/site-packages/transformers/optimization.py:306: FutureWarning: This implementation of AdamW is deprecated and will be removed in a future version. Use the PyTorch implementation torch.optim.AdamW instead, or set `no_deprecation_warning=True` to disable this warning\n",
+      "  warnings.warn(\n",
+      "***** Running training *****\n",
+      "  Num examples = 1995\n",
+      "  Num Epochs = 3\n",
+      "  Instantaneous batch size per device = 8\n",
+      "  Total train batch size (w. parallel, distributed & accumulation) = 8\n",
+      "  Gradient Accumulation steps = 1\n",
+      "  Total optimization steps = 750\n"
+     ]
+    },
+    {
+     "data": {
+      "text/plain": "<IPython.core.display.HTML object>",
+      "text/html": "\n    <div>\n      \n      <progress value='2' max='750' style='width:300px; height:20px; vertical-align: middle;'></progress>\n      [  2/750 : < :, Epoch 0.00/3]\n    </div>\n    <table border=\"1\" class=\"dataframe\">\n  <thead>\n <tr style=\"text-align: left;\">\n      <th>Step</th>\n      <th>Training Loss</th>\n    </tr>\n  </thead>\n  <tbody>\n  </tbody>\n</table><p>"
+     },
+     "metadata": {},
+     "output_type": "display_data"
+    },
+    {
+     "name": "stderr",
+     "output_type": "stream",
+     "text": [
+      "Saving model checkpoint to data/roberta-finetuned-subjqa/checkpoint-250\n",
+      "Configuration saved in data/roberta-finetuned-subjqa/checkpoint-250/config.json\n",
+      "Model weights saved in data/roberta-finetuned-subjqa/checkpoint-250/pytorch_model.bin\n",
+      "tokenizer config file saved in data/roberta-finetuned-subjqa/checkpoint-250/tokenizer_config.json\n",
+      "Special tokens file saved in data/roberta-finetuned-subjqa/checkpoint-250/special_tokens_map.json\n",
+      "Saving model checkpoint to data/roberta-finetuned-subjqa/checkpoint-500\n",
+      "Configuration saved in data/roberta-finetuned-subjqa/checkpoint-500/config.json\n",
+      "Model weights saved in data/roberta-finetuned-subjqa/checkpoint-500/pytorch_model.bin\n",
+      "tokenizer config file saved in data/roberta-finetuned-subjqa/checkpoint-500/tokenizer_config.json\n",
+      "Special tokens file saved in data/roberta-finetuned-subjqa/checkpoint-500/special_tokens_map.json\n",
+      "Saving model checkpoint to data/roberta-finetuned-subjqa/checkpoint-750\n",
+      "Configuration saved in data/roberta-finetuned-subjqa/checkpoint-750/config.json\n",
+      "Model weights saved in data/roberta-finetuned-subjqa/checkpoint-750/pytorch_model.bin\n",
+      "tokenizer config file saved in data/roberta-finetuned-subjqa/checkpoint-750/tokenizer_config.json\n",
+      "Special tokens file saved in data/roberta-finetuned-subjqa/checkpoint-750/special_tokens_map.json\n",
+      "\n",
+      "\n",
+      "Training completed. Do not forget to share your model on huggingface.co/models =)\n",
+      "\n",
+      "\n"
+     ]
+    },
+    {
+     "data": {
+      "text/plain": "TrainOutput(global_step=750, training_loss=0.4630465749104818, metrics={'train_runtime': 337.0904, 'train_samples_per_second': 17.755, 'train_steps_per_second': 2.225, 'total_flos': 1172895816798720.0, 'train_loss': 0.4630465749104818, 'epoch': 3.0})"
+     },
+     "execution_count": 57,
+     "metadata": {},
+     "output_type": "execute_result"
+    }
+   ],
+   "source": [
+    "args = TrainingArguments(\n",
+    "    output_dir=\"data/roberta-finetuned-subjqa\",\n",
+    "    overwrite_output_dir=True,\n",
+    "    evaluation_strategy=\"no\",\n",
+    "    save_strategy=\"epoch\",\n",
+    "    learning_rate=2e-5,\n",
+    "    num_train_epochs=3,\n",
+    "    weight_decay=0.01,\n",
+    "    fp16=True,\n",
+    ")\n",
+    "trainer = Trainer(\n",
+    "    model=model,\n",
+    "    args=args,\n",
+    "    train_dataset=roberta_train_dataset,\n",
+    "    eval_dataset=roberta_validation_dataset,\n",
+    "    tokenizer=tokenizer,\n",
+    ")\n",
+    "trainer.train()"
+   ],
+   "metadata": {
+    "collapsed": false,
+    "pycharm": {
+     "name": "#%%\n"
+    }
+   }
+  },
+  {
+   "cell_type": "code",
+   "execution_count": 58,
+   "outputs": [
+    {
+     "name": "stderr",
+     "output_type": "stream",
+     "text": [
+      "The following columns in the test set don't have a corresponding argument in `RobertaForQuestionAnswering.forward` and have been ignored: example_id, offset_mapping. If example_id, offset_mapping are not expected by `RobertaForQuestionAnswering.forward`,  you can safely ignore this message.\n",
+      "***** Running Prediction *****\n",
+      "  Num examples = 321\n",
+      "  Batch size = 8\n"
+     ]
+    },
+    {
+     "data": {
+      "text/plain": "<IPython.core.display.HTML object>",
+      "text/html": "\n    <div>\n      \n      <progress value='1' max='41' style='width:300px; height:20px; vertical-align: middle;'></progress>\n      [ 1/41 : < :]\n    </div>\n    "
+     },
+     "metadata": {},
+     "output_type": "display_data"
+    },
+    {
+     "data": {
+      "text/plain": "  0%|          | 0/265 [00:00<?, ?it/s]",
+      "application/vnd.jupyter.widget-view+json": {
+       "version_major": 2,
+       "version_minor": 0,
+       "model_id": "f5095780d2c947c9ad51f61f804a3484"
+      },
+      "application/json": {
+       "n": 0,
+       "total": 265,
+       "elapsed": 0.009457588195800781,
+       "ncols": null,
+       "nrows": 12,
+       "prefix": "",
+       "ascii": false,
+       "unit": "it",
+       "unit_scale": false,
+       "rate": null,
+       "bar_format": null,
+       "postfix": null,
+       "unit_divisor": 1000,
+       "initial": 0,
+       "colour": null
+      }
+     },
+     "metadata": {},
+     "output_type": "display_data"
+    },
+    {
+     "name": "stdout",
+     "output_type": "stream",
+     "text": [
+      "QUESTION:\t How was the hotel?\n",
+      "PREDICTED: No complaints\n",
+      "ACTUAL: ['excellent hotels']\n",
+      "QUESTION:\t How is the hotel?\n",
+      "PREDICTED: The hotel location was great\n",
+      "ACTUAL: ['The hotel location was great']\n",
+      "QUESTION:\t Is it value for money?\n",
+      "PREDICTED: excellent value for money\n",
+      "ACTUAL: ['And very reasonably priced. Overall, excellent value for money and highly recommended.']\n",
+      "{'exact_match': 29.81132075471698, 'f1': 46.72268395172683}\n"
+     ]
+    }
+   ],
+   "source": [
+    "predictions, _, _ = trainer.predict(roberta_validation_dataset)\n",
+    "start_logits, end_logits = predictions\n",
+    "_=compute_metrics(start_logits, end_logits, roberta_validation_dataset, subjqa[\"validation\"])"
+   ],
+   "metadata": {
+    "collapsed": false,
+    "pycharm": {
+     "name": "#%%\n"
+    }
+   }
+  }
+ ],
+ "metadata": {
+  "kernelspec": {
+   "display_name": "Python 3",
+   "language": "python",
+   "name": "python3"
+  },
+  "language_info": {
+   "codemirror_mode": {
+    "name": "ipython",
+    "version": 2
+   },
+   "file_extension": ".py",
+   "mimetype": "text/x-python",
+   "name": "python",
+   "nbconvert_exporter": "python",
+   "pygments_lexer": "ipython2",
+   "version": "2.7.6"
+  }
+ },
+ "nbformat": 4,
+ "nbformat_minor": 0
+}
\ No newline at end of file
diff --git a/Konopka_Raport.docx b/Konopka_Raport.docx
new file mode 100644
index 0000000..7b728f9
Binary files /dev/null and b/Konopka_Raport.docx differ
diff --git a/README.md b/README.md
new file mode 100644
index 0000000..bb6aa4e
--- /dev/null
+++ b/README.md
@@ -0,0 +1,42 @@
+# Projekt - Metody uczenia maszynowego w przetwarzaniu języka naturalnego
+
+## Wymagania
+
+Do oceny projektu proszę przygotować:
+
+- odnośnik do repozytorium z plikami projektu (jeśli istnieje)
+- demonstrację projektu
+- krótki raport, zawierający:
+    - cel projektu / definicję problemu (jakie zagadnienie Państwo rozwiązywali)
+    - opis użytych danych (w jaki sposób zebrano dane, czy i jak dokonano wstępnego przetworzenia danych, ile przykładów
+      zawierają zbiory uczący i testowy)
+    - opis wykorzystanych metod (jakich modeli Państwo użyli)
+    - tabelkę z wynikami ewaluacji
+    - ewentualne wnioski
+
+Wzór raportu znajduje się w pliku Wzór raportu.docx
+
+## Cel
+
+Celem projektu było stworzenie modelu, który odpowiada na subiektywne pytania na podstawie kontekstu.
+
+## Zbiór danych
+
+Dane pobrane są z przygotowanego korpusu [SubjQA](https://huggingface.co/datasets/subjqa) w
+wersji tripadvisor oraz restaurants. Jest to zbiór zawierający subiektywne pytania, takich jak "How do you like the soup?".  
+Zbiór danych jest zbudowany z tekstów ze stron internetowych, zawierających recenzje hoteli oraz resturacji oraz pytań i odpowiedzi.
+Nie wszystkie pytania mają odpowiedzi.
+Zbiór jest podzielony na train,test i validation, gdzie:
+  - tripadisor zawiera train:1165, test:230, validation:512 rekordów.
+  - restaurants zawiera train:1400, test:267, validation:266 rekordów.
+
+Ponieważ przy trenowaniu nie ma potrzeby używać splitu test, więc po połączeniu datasetów, przyłączeniu
+test do train i usunięciu rekordów ktore nie mają odpowiedzi na pytanie powstaje zbiór o wielkości:
+ - train: 1666
+ - validation: 265
+
+Tokenizowany kontekst musi mieć ograniczoną wielkość, w tym przypadku 384, 
+z tego powodu dany zbiór został przetworzony używając stride=128 i max_lenght=384, co sprawiło że całkowita 
+wielkość zbiru wyniosła:
+ - train 2030 features
+ - validation 327 features
\ No newline at end of file
diff --git a/requirements.txt b/requirements.txt
new file mode 100644
index 0000000..d51fef4
--- /dev/null
+++ b/requirements.txt
@@ -0,0 +1,10 @@
+jupyter
+torch==1.12.1+cu116
+torchaudio==0.12.1+cu116
+torchvision==0.13.1+cu116
+--extra-index-url https://download.pytorch.org/whl/cu116
+transformers==4.21.1
+datasets==2.4.0
+matplotlib
+pandas
+evaluate
\ No newline at end of file