01_prompt_testing.ipynb

{
 "cells": [
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "# Prompt selection and testing\n",
    "\n",
    "This notebook chooses the most appropriate prompt and prompt structure for the OCR correction. "
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 58,
   "metadata": {},
   "outputs": [],
   "source": [
    "#import config  # Import your config.py file this contains you openai api key\n",
    "import pandas as pd\n",
    "import numpy as np\n",
    "import os\n",
    "from llm_comparison_toolkit import RateLimiter, get_response_openai, get_response_anthropic,  create_config_dict_func, compare_request_configurations\n",
    "from evaluate import load\n",
    "from evaluation_funcs import evaluate_correction_performance, evaluate_correction_performance_folders, get_metric_error_reduction\n",
    "from helper_functions import files_to_df_func, files_to_df_core_func\n",
    "import seaborn as sns\n",
    "import matplotlib.pyplot as plt\n",
    "from dotenv import load_dotenv, find_dotenv\n",
    "load_dotenv()\n",
    "save_appendix = os.getenv(\"save_appendix\")\n",
    "\n",
    "dev_data_folder = 'data/dev_data'\n",
    "dev_transcripts = os.path.join(dev_data_folder, 'dev_data_transcript')\n",
    "dev_raw_ocr_folder =  os.path.join(dev_data_folder,'dev_raw_ocr' )\n",
    "dev_system_message_folder = os.path.join(dev_data_folder,'dev_system_message_variants' )\n",
    "\n",
    "#load the dev and test sets for prompt development and selection\n",
    "dev_data_df = pd.read_csv(os.path.join(dev_data_folder,'dev_data_raw.csv'))\n",
    "\n",
    "\n",
    "#for saving data to be used in the analysis\n",
    "if not os.path.exists('data/analysis'):\n",
    "    os.makedirs('data/analysis')"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "# Explore different system prompts\n",
    "\n",
    "This explores a range of system prompt to find the one that appears to work the best, we use gpt4 as the baseline model.\n",
    "\n",
    "Although there is no comparison with all models we do test gpt3.5 gpt4, clause haiku and claude opus, in addition we put the prompt in the system message and the prompt after the text becuase this may affect the quality of the result. When the prompt is after the text the response has \"nosm_\" (no system message) appended to the file name"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 2,
   "metadata": {},
   "outputs": [],
   "source": [
    "#Create a modular set of system messages that can be combined in different ways\n",
    "basic_prompt = \"Please recover the text from the corrupted OCR.\"\n",
    "expertise_prompt = \"You are an expert in post-OCR correction of documents.\"\n",
    "recover_prompt = \"Using the context available from the text please recover the most likely original text from the corrupted OCR.\"\n",
    "publication_context_prompt = \"The text is from an english newspaper in the 1800's.\"\n",
    "text_context_prompt = \"The text may be an advert or article and may be missing the beggining or end.\"\n",
    "additional_instructions_prompt = \"Do not add any text, commentary, or lead in sentences beyond the recovered text. Do not add a title, or any introductions.\"\n",
    "\n",
    "#combine all the message parts into a variety of system messages, a tuple is used where 0 is the name of the message and 1 is the message itself\n",
    "#N.B. This is not and exhaustive combination as that would be very expensive and likley not yield significantly better results\n",
    "system_messages_list = [\n",
    "('basic_prompt', basic_prompt),\n",
    "('expert_basic_prompt', expertise_prompt + ' '+ basic_prompt),\n",
    "('expert_recover_prompt', expertise_prompt + ' '+ recover_prompt),\n",
    "('expert_recover_publication_prompt', expertise_prompt + ' '+ recover_prompt + ' ' + publication_context_prompt),\n",
    "('expert_recover_text_prompt', expertise_prompt + ' '+ recover_prompt + ' ' + text_context_prompt),\n",
    "('expert_recover_publication_text_prompt', expertise_prompt + ' '+ recover_prompt + ' ' + publication_context_prompt + ' ' + text_context_prompt),\n",
    "('expert_recover_instructions_prompt', expertise_prompt + ' '+ recover_prompt + ' ' + additional_instructions_prompt),\n",
    "('full_context_prompt', expertise_prompt + ' '+ recover_prompt + ' ' + publication_context_prompt + ' ' + text_context_prompt+ ' ' + additional_instructions_prompt)\n",
    "]"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "The below function is used to make the creation of the config dictionaries for the test more compact and increase readability"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 3,
   "metadata": {},
   "outputs": [],
   "source": [
    "def create_message_test_configs(system_messages_list, get_response_func, engine):\n",
    "    message_test_configs = []\n",
    "    for iter_system_message in system_messages_list:\n",
    "        message_test_configs.append(\n",
    "            create_config_dict_func(\n",
    "                get_response_func=get_response_func,\n",
    "                rate_limiter=RateLimiter(50000),\n",
    "                engine=engine,\n",
    "                system_message_template=iter_system_message[1],\n",
    "                prompt_template=\"{content_html}\",\n",
    "                additional_args={'response_name': iter_system_message[0]}\n",
    "            )\n",
    "        )\n",
    "        message_test_configs.append(\n",
    "            create_config_dict_func(\n",
    "                get_response_func=get_response_func,\n",
    "                rate_limiter=RateLimiter(50000),\n",
    "                engine=engine,\n",
    "                system_message_template=\"\",\n",
    "                prompt_template=\"{content_html}\" + f\"\"\"\\n\\n\"\"\" + iter_system_message[1],\n",
    "                additional_args={'response_name': \"nosm_\"+iter_system_message[0]}\n",
    "            )\n",
    "        )\n",
    "    return message_test_configs"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## Create configs and run tests"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 4,
   "metadata": {},
   "outputs": [],
   "source": [
    "#gpt configs\n",
    "gpt3_prompt_testing_configs = create_message_test_configs(system_messages_list, get_response_openai, \"gpt-3.5-turbo\")\n",
    "gpt4_prompt_testing_configs = create_message_test_configs(system_messages_list, get_response_openai, 'gpt-4-turbo-preview')\n",
    "\n",
    "#claude configs\n",
    "haiku_prompt_testing_configs = create_message_test_configs(system_messages_list, get_response_anthropic, \"claude-3-haiku-20240307\")\n",
    "opus_prompt_testing_configs = create_message_test_configs(system_messages_list, get_response_anthropic, \"claude-3-opus-20240229\")\n",
    "\n",
    "#run the experiment on all the prompt configs and save to the folder\n",
    "compare_request_configurations(dev_data_df, \n",
    "                               gpt3_prompt_testing_configs + gpt4_prompt_testing_configs + haiku_prompt_testing_configs + opus_prompt_testing_configs,\n",
    "                               folder_path = dev_system_message_folder )"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 5,
   "metadata": {},
   "outputs": [],
   "source": [
    "\n",
    "prompt_test_df = []\n",
    "for folder in os.listdir(dev_system_message_folder):\n",
    "    df = files_to_df_core_func(dev_system_message_folder)\n",
    "    df['folder'] = folder\n",
    "    prompt_test_df.append(df)\n",
    "\n",
    "prompt_test_df = pd.concat(prompt_test_df, ignore_index=True)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "# Evaluate system prompt tests\n",
    "\n",
    "We evaluate the system prompts below to see if thre is any significant difference between the prompts\n"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 6,
   "metadata": {},
   "outputs": [],
   "source": [
    "wer = load(\"wer\")\n",
    "cer = load(\"cer\")"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 7,
   "metadata": {},
   "outputs": [],
   "source": [
    "raw_dev_ocr_scores = evaluate_correction_performance(dev_raw_ocr_folder , dev_transcripts, wer, cer, 'raw_ocr')"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 8,
   "metadata": {},
   "outputs": [],
   "source": [
    "corrected_folder = dev_system_message_folder\n",
    "\n",
    "performance_eval = evaluate_correction_performance_folders(corrected_folder, dev_transcripts, wer, cer)\n",
    "\n",
    "performance_eval =  performance_eval.loc[(performance_eval['File Name']!='slug_ar02501_periodical_pc_issue_tec_06121884_page_number_25.txt'),:]\n",
    "\n",
    "performance_eval['type'] = performance_eval['type'].str.replace(\"claude-3-haiku-20240307\", \"haiku\").replace(\"gpt-3.5-turbo\", \"gpt-3.5\")\n",
    "\n",
    "performance_eval['model'] = performance_eval['type'].str.split('_').str[-1]"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 9,
   "metadata": {},
   "outputs": [],
   "source": [
    "test = get_metric_error_reduction(performance_eval, raw_dev_ocr_scores)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 10,
   "metadata": {},
   "outputs": [
    {
     "data": {
      "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 tr th {\n",
       "        text-align: left;\n",
       "    }\n",
       "\n",
       "    .dataframe thead tr:last-of-type th {\n",
       "        text-align: right;\n",
       "    }\n",
       "</style>\n",
       "<table border=\"1\" class=\"dataframe\">\n",
       "  <thead>\n",
       "    <tr>\n",
       "      <th></th>\n",
       "      <th colspan=\"2\" halign=\"left\">WER</th>\n",
       "      <th colspan=\"2\" halign=\"left\">CER</th>\n",
       "      <th colspan=\"2\" halign=\"left\">lev_dist</th>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th></th>\n",
       "      <th>mean</th>\n",
       "      <th>50%</th>\n",
       "      <th>mean</th>\n",
       "      <th>50%</th>\n",
       "      <th>mean</th>\n",
       "      <th>50%</th>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>type</th>\n",
       "      <th></th>\n",
       "      <th></th>\n",
       "      <th></th>\n",
       "      <th></th>\n",
       "      <th></th>\n",
       "      <th></th>\n",
       "    </tr>\n",
       "  </thead>\n",
       "  <tbody>\n",
       "    <tr>\n",
       "      <th>expert_recover_text_prompt_haiku</th>\n",
       "      <td>-4.66</td>\n",
       "      <td>8.18</td>\n",
       "      <td>-296.51</td>\n",
       "      <td>-197.56</td>\n",
       "      <td>-283.62</td>\n",
       "      <td>-181.11</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>nosm_expert_recover_publication_text_prompt_haiku</th>\n",
       "      <td>2.83</td>\n",
       "      <td>12.73</td>\n",
       "      <td>-294.84</td>\n",
       "      <td>-152.04</td>\n",
       "      <td>-281.51</td>\n",
       "      <td>-147.84</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>expert_recover_publication_text_prompt_haiku</th>\n",
       "      <td>16.26</td>\n",
       "      <td>31.82</td>\n",
       "      <td>-209.72</td>\n",
       "      <td>-115.22</td>\n",
       "      <td>-200.48</td>\n",
       "      <td>-110.20</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>nosm_expert_recover_text_prompt_haiku</th>\n",
       "      <td>0.41</td>\n",
       "      <td>7.50</td>\n",
       "      <td>-297.10</td>\n",
       "      <td>-106.33</td>\n",
       "      <td>-283.08</td>\n",
       "      <td>-103.45</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>expert_recover_prompt_haiku</th>\n",
       "      <td>37.93</td>\n",
       "      <td>45.61</td>\n",
       "      <td>-90.47</td>\n",
       "      <td>-38.91</td>\n",
       "      <td>-86.93</td>\n",
       "      <td>-40.09</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>...</th>\n",
       "      <td>...</td>\n",
       "      <td>...</td>\n",
       "      <td>...</td>\n",
       "      <td>...</td>\n",
       "      <td>...</td>\n",
       "      <td>...</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>nosm_full_context_prompt_gpt-4-turbo-preview</th>\n",
       "      <td>73.21</td>\n",
       "      <td>77.19</td>\n",
       "      <td>54.38</td>\n",
       "      <td>63.78</td>\n",
       "      <td>51.48</td>\n",
       "      <td>62.23</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>nosm_expert_recover_pub_instructions_prompt_claude-3-opus-20240229</th>\n",
       "      <td>71.88</td>\n",
       "      <td>78.71</td>\n",
       "      <td>40.15</td>\n",
       "      <td>68.74</td>\n",
       "      <td>37.37</td>\n",
       "      <td>62.84</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>nosm_full_context_prompt_claude-3-opus-20240229</th>\n",
       "      <td>72.74</td>\n",
       "      <td>80.46</td>\n",
       "      <td>58.40</td>\n",
       "      <td>68.74</td>\n",
       "      <td>54.51</td>\n",
       "      <td>62.84</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>expert_recover_instructions_prompt_claude-3-opus-20240229</th>\n",
       "      <td>72.91</td>\n",
       "      <td>81.03</td>\n",
       "      <td>46.22</td>\n",
       "      <td>65.41</td>\n",
       "      <td>42.21</td>\n",
       "      <td>63.30</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>nosm_expert_recover_prompt_gpt-4-turbo-preview</th>\n",
       "      <td>72.10</td>\n",
       "      <td>75.86</td>\n",
       "      <td>49.29</td>\n",
       "      <td>65.55</td>\n",
       "      <td>46.21</td>\n",
       "      <td>63.79</td>\n",
       "    </tr>\n",
       "  </tbody>\n",
       "</table>\n",
       "<p>72 rows × 6 columns</p>\n",
       "</div>"
      ],
      "text/plain": [
       "                                                      WER            CER  \\\n",
       "                                                     mean    50%    mean   \n",
       "type                                                                       \n",
       "expert_recover_text_prompt_haiku                    -4.66   8.18 -296.51   \n",
       "nosm_expert_recover_publication_text_prompt_haiku    2.83  12.73 -294.84   \n",
       "expert_recover_publication_text_prompt_haiku        16.26  31.82 -209.72   \n",
       "nosm_expert_recover_text_prompt_haiku                0.41   7.50 -297.10   \n",
       "expert_recover_prompt_haiku                         37.93  45.61  -90.47   \n",
       "...                                                   ...    ...     ...   \n",
       "nosm_full_context_prompt_gpt-4-turbo-preview        73.21  77.19   54.38   \n",
       "nosm_expert_recover_pub_instructions_prompt_cla...  71.88  78.71   40.15   \n",
       "nosm_full_context_prompt_claude-3-opus-20240229     72.74  80.46   58.40   \n",
       "expert_recover_instructions_prompt_claude-3-opu...  72.91  81.03   46.22   \n",
       "nosm_expert_recover_prompt_gpt-4-turbo-preview      72.10  75.86   49.29   \n",
       "\n",
       "                                                           lev_dist          \n",
       "                                                       50%     mean     50%  \n",
       "type                                                                         \n",
       "expert_recover_text_prompt_haiku                   -197.56  -283.62 -181.11  \n",
       "nosm_expert_recover_publication_text_prompt_haiku  -152.04  -281.51 -147.84  \n",
       "expert_recover_publication_text_prompt_haiku       -115.22  -200.48 -110.20  \n",
       "nosm_expert_recover_text_prompt_haiku              -106.33  -283.08 -103.45  \n",
       "expert_recover_prompt_haiku                         -38.91   -86.93  -40.09  \n",
       "...                                                    ...      ...     ...  \n",
       "nosm_full_context_prompt_gpt-4-turbo-preview         63.78    51.48   62.23  \n",
       "nosm_expert_recover_pub_instructions_prompt_cla...   68.74    37.37   62.84  \n",
       "nosm_full_context_prompt_claude-3-opus-20240229      68.74    54.51   62.84  \n",
       "expert_recover_instructions_prompt_claude-3-opu...   65.41    42.21   63.30  \n",
       "nosm_expert_recover_prompt_gpt-4-turbo-preview       65.55    46.21   63.79  \n",
       "\n",
       "[72 rows x 6 columns]"
      ]
     },
     "execution_count": 10,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "test.groupby('type').describe().filter(regex = '50|mean').round(2).sort_values(('lev_dist', '50%'))"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": []
  },
  {
   "cell_type": "code",
   "execution_count": 11,
   "metadata": {},
   "outputs": [
    {
     "data": {
      "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 tr th {\n",
       "        text-align: left;\n",
       "    }\n",
       "\n",
       "    .dataframe thead tr:last-of-type th {\n",
       "        text-align: right;\n",
       "    }\n",
       "</style>\n",
       "<table border=\"1\" class=\"dataframe\">\n",
       "  <thead>\n",
       "    <tr>\n",
       "      <th></th>\n",
       "      <th colspan=\"2\" halign=\"left\">WER</th>\n",
       "      <th colspan=\"2\" halign=\"left\">CER</th>\n",
       "      <th colspan=\"2\" halign=\"left\">lev_dist</th>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th></th>\n",
       "      <th>mean</th>\n",
       "      <th>50%</th>\n",
       "      <th>mean</th>\n",
       "      <th>50%</th>\n",
       "      <th>mean</th>\n",
       "      <th>50%</th>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>type</th>\n",
       "      <th></th>\n",
       "      <th></th>\n",
       "      <th></th>\n",
       "      <th></th>\n",
       "      <th></th>\n",
       "      <th></th>\n",
       "    </tr>\n",
       "  </thead>\n",
       "  <tbody>\n",
       "    <tr>\n",
       "      <th>nosm_expert_recover_publication_text_prompt_claude-3-opus-20240229</th>\n",
       "      <td>55.27</td>\n",
       "      <td>74.37</td>\n",
       "      <td>-30.47</td>\n",
       "      <td>43.74</td>\n",
       "      <td>-33.26</td>\n",
       "      <td>40.25</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>nosm_expert_recover_publication_prompt_claude-3-opus-20240229</th>\n",
       "      <td>71.07</td>\n",
       "      <td>73.90</td>\n",
       "      <td>44.04</td>\n",
       "      <td>50.23</td>\n",
       "      <td>39.63</td>\n",
       "      <td>45.26</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>nosm_expert_recover_text_prompt_claude-3-opus-20240229</th>\n",
       "      <td>70.36</td>\n",
       "      <td>76.02</td>\n",
       "      <td>49.85</td>\n",
       "      <td>57.28</td>\n",
       "      <td>45.44</td>\n",
       "      <td>46.78</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>nosm_basic_prompt_claude-3-opus-20240229</th>\n",
       "      <td>69.83</td>\n",
       "      <td>72.41</td>\n",
       "      <td>46.73</td>\n",
       "      <td>52.74</td>\n",
       "      <td>42.14</td>\n",
       "      <td>47.06</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>nosm_expert_recover_prompt_claude-3-opus-20240229</th>\n",
       "      <td>72.06</td>\n",
       "      <td>76.88</td>\n",
       "      <td>48.80</td>\n",
       "      <td>53.71</td>\n",
       "      <td>44.31</td>\n",
       "      <td>51.41</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>expert_recover_prompt_claude-3-opus-20240229</th>\n",
       "      <td>62.40</td>\n",
       "      <td>73.90</td>\n",
       "      <td>4.22</td>\n",
       "      <td>56.11</td>\n",
       "      <td>2.23</td>\n",
       "      <td>51.65</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>expert_recover_pub_instructions_prompt_claude-3-opus-20240229</th>\n",
       "      <td>74.58</td>\n",
       "      <td>78.89</td>\n",
       "      <td>57.67</td>\n",
       "      <td>66.22</td>\n",
       "      <td>52.95</td>\n",
       "      <td>51.72</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>expert_recover_publication_prompt_claude-3-opus-20240229</th>\n",
       "      <td>70.00</td>\n",
       "      <td>74.37</td>\n",
       "      <td>41.13</td>\n",
       "      <td>55.14</td>\n",
       "      <td>36.57</td>\n",
       "      <td>51.72</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>expert_recover_text_prompt_claude-3-opus-20240229</th>\n",
       "      <td>71.19</td>\n",
       "      <td>72.41</td>\n",
       "      <td>43.40</td>\n",
       "      <td>56.22</td>\n",
       "      <td>38.97</td>\n",
       "      <td>51.72</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>expert_recover_publication_text_prompt_claude-3-opus-20240229</th>\n",
       "      <td>69.84</td>\n",
       "      <td>74.14</td>\n",
       "      <td>41.57</td>\n",
       "      <td>57.01</td>\n",
       "      <td>37.36</td>\n",
       "      <td>53.19</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>full_context_prompt_claude-3-opus-20240229</th>\n",
       "      <td>67.83</td>\n",
       "      <td>78.89</td>\n",
       "      <td>24.82</td>\n",
       "      <td>56.56</td>\n",
       "      <td>22.79</td>\n",
       "      <td>53.29</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>basic_prompt_claude-3-opus-20240229</th>\n",
       "      <td>71.42</td>\n",
       "      <td>75.86</td>\n",
       "      <td>41.21</td>\n",
       "      <td>57.01</td>\n",
       "      <td>37.20</td>\n",
       "      <td>53.62</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>nosm_expert_basic_prompt_claude-3-opus-20240229</th>\n",
       "      <td>72.86</td>\n",
       "      <td>77.59</td>\n",
       "      <td>55.84</td>\n",
       "      <td>61.46</td>\n",
       "      <td>50.93</td>\n",
       "      <td>56.68</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>expert_basic_prompt_claude-3-opus-20240229</th>\n",
       "      <td>75.14</td>\n",
       "      <td>79.57</td>\n",
       "      <td>58.65</td>\n",
       "      <td>64.39</td>\n",
       "      <td>53.70</td>\n",
       "      <td>59.45</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>nosm_expert_recover_instructions_prompt_claude-3-opus-20240229</th>\n",
       "      <td>73.45</td>\n",
       "      <td>81.03</td>\n",
       "      <td>55.15</td>\n",
       "      <td>66.83</td>\n",
       "      <td>50.36</td>\n",
       "      <td>61.12</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>nosm_expert_recover_pub_instructions_prompt_claude-3-opus-20240229</th>\n",
       "      <td>71.88</td>\n",
       "      <td>78.71</td>\n",
       "      <td>40.15</td>\n",
       "      <td>68.74</td>\n",
       "      <td>37.37</td>\n",
       "      <td>62.84</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>nosm_full_context_prompt_claude-3-opus-20240229</th>\n",
       "      <td>72.74</td>\n",
       "      <td>80.46</td>\n",
       "      <td>58.40</td>\n",
       "      <td>68.74</td>\n",
       "      <td>54.51</td>\n",
       "      <td>62.84</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>expert_recover_instructions_prompt_claude-3-opus-20240229</th>\n",
       "      <td>72.91</td>\n",
       "      <td>81.03</td>\n",
       "      <td>46.22</td>\n",
       "      <td>65.41</td>\n",
       "      <td>42.21</td>\n",
       "      <td>63.30</td>\n",
       "    </tr>\n",
       "  </tbody>\n",
       "</table>\n",
       "</div>"
      ],
      "text/plain": [
       "                                                      WER           CER  \\\n",
       "                                                     mean    50%   mean   \n",
       "type                                                                      \n",
       "nosm_expert_recover_publication_text_prompt_cla...  55.27  74.37 -30.47   \n",
       "nosm_expert_recover_publication_prompt_claude-3...  71.07  73.90  44.04   \n",
       "nosm_expert_recover_text_prompt_claude-3-opus-2...  70.36  76.02  49.85   \n",
       "nosm_basic_prompt_claude-3-opus-20240229            69.83  72.41  46.73   \n",
       "nosm_expert_recover_prompt_claude-3-opus-20240229   72.06  76.88  48.80   \n",
       "expert_recover_prompt_claude-3-opus-20240229        62.40  73.90   4.22   \n",
       "expert_recover_pub_instructions_prompt_claude-3...  74.58  78.89  57.67   \n",
       "expert_recover_publication_prompt_claude-3-opus...  70.00  74.37  41.13   \n",
       "expert_recover_text_prompt_claude-3-opus-20240229   71.19  72.41  43.40   \n",
       "expert_recover_publication_text_prompt_claude-3...  69.84  74.14  41.57   \n",
       "full_context_prompt_claude-3-opus-20240229          67.83  78.89  24.82   \n",
       "basic_prompt_claude-3-opus-20240229                 71.42  75.86  41.21   \n",
       "nosm_expert_basic_prompt_claude-3-opus-20240229     72.86  77.59  55.84   \n",
       "expert_basic_prompt_claude-3-opus-20240229          75.14  79.57  58.65   \n",
       "nosm_expert_recover_instructions_prompt_claude-...  73.45  81.03  55.15   \n",
       "nosm_expert_recover_pub_instructions_prompt_cla...  71.88  78.71  40.15   \n",
       "nosm_full_context_prompt_claude-3-opus-20240229     72.74  80.46  58.40   \n",
       "expert_recover_instructions_prompt_claude-3-opu...  72.91  81.03  46.22   \n",
       "\n",
       "                                                          lev_dist         \n",
       "                                                      50%     mean    50%  \n",
       "type                                                                       \n",
       "nosm_expert_recover_publication_text_prompt_cla...  43.74   -33.26  40.25  \n",
       "nosm_expert_recover_publication_prompt_claude-3...  50.23    39.63  45.26  \n",
       "nosm_expert_recover_text_prompt_claude-3-opus-2...  57.28    45.44  46.78  \n",
       "nosm_basic_prompt_claude-3-opus-20240229            52.74    42.14  47.06  \n",
       "nosm_expert_recover_prompt_claude-3-opus-20240229   53.71    44.31  51.41  \n",
       "expert_recover_prompt_claude-3-opus-20240229        56.11     2.23  51.65  \n",
       "expert_recover_pub_instructions_prompt_claude-3...  66.22    52.95  51.72  \n",
       "expert_recover_publication_prompt_claude-3-opus...  55.14    36.57  51.72  \n",
       "expert_recover_text_prompt_claude-3-opus-20240229   56.22    38.97  51.72  \n",
       "expert_recover_publication_text_prompt_claude-3...  57.01    37.36  53.19  \n",
       "full_context_prompt_claude-3-opus-20240229          56.56    22.79  53.29  \n",
       "basic_prompt_claude-3-opus-20240229                 57.01    37.20  53.62  \n",
       "nosm_expert_basic_prompt_claude-3-opus-20240229     61.46    50.93  56.68  \n",
       "expert_basic_prompt_claude-3-opus-20240229          64.39    53.70  59.45  \n",
       "nosm_expert_recover_instructions_prompt_claude-...  66.83    50.36  61.12  \n",
       "nosm_expert_recover_pub_instructions_prompt_cla...  68.74    37.37  62.84  \n",
       "nosm_full_context_prompt_claude-3-opus-20240229     68.74    54.51  62.84  \n",
       "expert_recover_instructions_prompt_claude-3-opu...  65.41    42.21  63.30  "
      ]
     },
     "execution_count": 11,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "test.loc[test['type'].str.contains('opus')].groupby('type').describe().filter(regex = '50|mean').round(2).sort_values(('lev_dist', '50%'))"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 17,
   "metadata": {},
   "outputs": [
    {
     "data": {
      "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 tr th {\n",
       "        text-align: left;\n",
       "    }\n",
       "\n",
       "    .dataframe thead tr:last-of-type th {\n",
       "        text-align: right;\n",
       "    }\n",
       "</style>\n",
       "<table border=\"1\" class=\"dataframe\">\n",
       "  <thead>\n",
       "    <tr>\n",
       "      <th></th>\n",
       "      <th></th>\n",
       "      <th colspan=\"2\" halign=\"left\">WER</th>\n",
       "      <th colspan=\"2\" halign=\"left\">CER</th>\n",
       "      <th colspan=\"2\" halign=\"left\">lev_dist</th>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th></th>\n",
       "      <th></th>\n",
       "      <th>mean</th>\n",
       "      <th>50%</th>\n",
       "      <th>mean</th>\n",
       "      <th>50%</th>\n",
       "      <th>mean</th>\n",
       "      <th>50%</th>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>type</th>\n",
       "      <th>model</th>\n",
       "      <th></th>\n",
       "      <th></th>\n",
       "      <th></th>\n",
       "      <th></th>\n",
       "      <th></th>\n",
       "      <th></th>\n",
       "    </tr>\n",
       "  </thead>\n",
       "  <tbody>\n",
       "    <tr>\n",
       "      <th>nosm_expert_recover_instructions_prompt_gpt-4-turbo-preview</th>\n",
       "      <th>gpt-4-turbo-preview</th>\n",
       "      <td>0.23</td>\n",
       "      <td>0.15</td>\n",
       "      <td>0.13</td>\n",
       "      <td>0.06</td>\n",
       "      <td>148.62</td>\n",
       "      <td>77.0</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>full_context_prompt_gpt-4-turbo-preview</th>\n",
       "      <th>gpt-4-turbo-preview</th>\n",
       "      <td>0.23</td>\n",
       "      <td>0.13</td>\n",
       "      <td>0.13</td>\n",
       "      <td>0.06</td>\n",
       "      <td>151.38</td>\n",
       "      <td>80.0</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>expert_recover_pub_instructions_prompt_gpt-4-turbo-preview</th>\n",
       "      <th>gpt-4-turbo-preview</th>\n",
       "      <td>0.23</td>\n",
       "      <td>0.14</td>\n",
       "      <td>0.13</td>\n",
       "      <td>0.05</td>\n",
       "      <td>153.43</td>\n",
       "      <td>80.0</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>expert_recover_instructions_prompt_claude-3-opus-20240229</th>\n",
       "      <th>claude-3-opus-20240229</th>\n",
       "      <td>0.24</td>\n",
       "      <td>0.19</td>\n",
       "      <td>0.14</td>\n",
       "      <td>0.05</td>\n",
       "      <td>183.24</td>\n",
       "      <td>84.0</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>nosm_full_context_prompt_gpt-4-turbo-preview</th>\n",
       "      <th>gpt-4-turbo-preview</th>\n",
       "      <td>0.24</td>\n",
       "      <td>0.16</td>\n",
       "      <td>0.14</td>\n",
       "      <td>0.07</td>\n",
       "      <td>160.43</td>\n",
       "      <td>87.0</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>...</th>\n",
       "      <th>...</th>\n",
       "      <td>...</td>\n",
       "      <td>...</td>\n",
       "      <td>...</td>\n",
       "      <td>...</td>\n",
       "      <td>...</td>\n",
       "      <td>...</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>nosm_expert_recover_publication_prompt_haiku</th>\n",
       "      <th>haiku</th>\n",
       "      <td>0.59</td>\n",
       "      <td>0.53</td>\n",
       "      <td>0.44</td>\n",
       "      <td>0.42</td>\n",
       "      <td>438.71</td>\n",
       "      <td>396.0</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>expert_recover_publication_text_prompt_haiku</th>\n",
       "      <th>haiku</th>\n",
       "      <td>0.72</td>\n",
       "      <td>0.51</td>\n",
       "      <td>0.58</td>\n",
       "      <td>0.41</td>\n",
       "      <td>581.00</td>\n",
       "      <td>526.0</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>expert_recover_text_prompt_haiku</th>\n",
       "      <th>haiku</th>\n",
       "      <td>0.84</td>\n",
       "      <td>0.90</td>\n",
       "      <td>0.68</td>\n",
       "      <td>0.70</td>\n",
       "      <td>785.76</td>\n",
       "      <td>543.0</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>nosm_expert_recover_text_prompt_haiku</th>\n",
       "      <th>haiku</th>\n",
       "      <td>0.82</td>\n",
       "      <td>0.81</td>\n",
       "      <td>0.64</td>\n",
       "      <td>0.65</td>\n",
       "      <td>630.95</td>\n",
       "      <td>549.0</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>nosm_expert_recover_publication_text_prompt_haiku</th>\n",
       "      <th>haiku</th>\n",
       "      <td>0.82</td>\n",
       "      <td>0.68</td>\n",
       "      <td>0.66</td>\n",
       "      <td>0.54</td>\n",
       "      <td>605.81</td>\n",
       "      <td>575.0</td>\n",
       "    </tr>\n",
       "  </tbody>\n",
       "</table>\n",
       "<p>72 rows × 6 columns</p>\n",
       "</div>"
      ],
      "text/plain": [
       "                                                                            WER  \\\n",
       "                                                                           mean   \n",
       "type                                               model                          \n",
       "nosm_expert_recover_instructions_prompt_gpt-4-t... gpt-4-turbo-preview     0.23   \n",
       "full_context_prompt_gpt-4-turbo-preview            gpt-4-turbo-preview     0.23   \n",
       "expert_recover_pub_instructions_prompt_gpt-4-tu... gpt-4-turbo-preview     0.23   \n",
       "expert_recover_instructions_prompt_claude-3-opu... claude-3-opus-20240229  0.24   \n",
       "nosm_full_context_prompt_gpt-4-turbo-preview       gpt-4-turbo-preview     0.24   \n",
       "...                                                                         ...   \n",
       "nosm_expert_recover_publication_prompt_haiku       haiku                   0.59   \n",
       "expert_recover_publication_text_prompt_haiku       haiku                   0.72   \n",
       "expert_recover_text_prompt_haiku                   haiku                   0.84   \n",
       "nosm_expert_recover_text_prompt_haiku              haiku                   0.82   \n",
       "nosm_expert_recover_publication_text_prompt_haiku  haiku                   0.82   \n",
       "\n",
       "                                                                                 \\\n",
       "                                                                            50%   \n",
       "type                                               model                          \n",
       "nosm_expert_recover_instructions_prompt_gpt-4-t... gpt-4-turbo-preview     0.15   \n",
       "full_context_prompt_gpt-4-turbo-preview            gpt-4-turbo-preview     0.13   \n",
       "expert_recover_pub_instructions_prompt_gpt-4-tu... gpt-4-turbo-preview     0.14   \n",
       "expert_recover_instructions_prompt_claude-3-opu... claude-3-opus-20240229  0.19   \n",
       "nosm_full_context_prompt_gpt-4-turbo-preview       gpt-4-turbo-preview     0.16   \n",
       "...                                                                         ...   \n",
       "nosm_expert_recover_publication_prompt_haiku       haiku                   0.53   \n",
       "expert_recover_publication_text_prompt_haiku       haiku                   0.51   \n",
       "expert_recover_text_prompt_haiku                   haiku                   0.90   \n",
       "nosm_expert_recover_text_prompt_haiku              haiku                   0.81   \n",
       "nosm_expert_recover_publication_text_prompt_haiku  haiku                   0.68   \n",
       "\n",
       "                                                                            CER  \\\n",
       "                                                                           mean   \n",
       "type                                               model                          \n",
       "nosm_expert_recover_instructions_prompt_gpt-4-t... gpt-4-turbo-preview     0.13   \n",
       "full_context_prompt_gpt-4-turbo-preview            gpt-4-turbo-preview     0.13   \n",
       "expert_recover_pub_instructions_prompt_gpt-4-tu... gpt-4-turbo-preview     0.13   \n",
       "expert_recover_instructions_prompt_claude-3-opu... claude-3-opus-20240229  0.14   \n",
       "nosm_full_context_prompt_gpt-4-turbo-preview       gpt-4-turbo-preview     0.14   \n",
       "...                                                                         ...   \n",
       "nosm_expert_recover_publication_prompt_haiku       haiku                   0.44   \n",
       "expert_recover_publication_text_prompt_haiku       haiku                   0.58   \n",
       "expert_recover_text_prompt_haiku                   haiku                   0.68   \n",
       "nosm_expert_recover_text_prompt_haiku              haiku                   0.64   \n",
       "nosm_expert_recover_publication_text_prompt_haiku  haiku                   0.66   \n",
       "\n",
       "                                                                                 \\\n",
       "                                                                            50%   \n",
       "type                                               model                          \n",
       "nosm_expert_recover_instructions_prompt_gpt-4-t... gpt-4-turbo-preview     0.06   \n",
       "full_context_prompt_gpt-4-turbo-preview            gpt-4-turbo-preview     0.06   \n",
       "expert_recover_pub_instructions_prompt_gpt-4-tu... gpt-4-turbo-preview     0.05   \n",
       "expert_recover_instructions_prompt_claude-3-opu... claude-3-opus-20240229  0.05   \n",
       "nosm_full_context_prompt_gpt-4-turbo-preview       gpt-4-turbo-preview     0.07   \n",
       "...                                                                         ...   \n",
       "nosm_expert_recover_publication_prompt_haiku       haiku                   0.42   \n",
       "expert_recover_publication_text_prompt_haiku       haiku                   0.41   \n",
       "expert_recover_text_prompt_haiku                   haiku                   0.70   \n",
       "nosm_expert_recover_text_prompt_haiku              haiku                   0.65   \n",
       "nosm_expert_recover_publication_text_prompt_haiku  haiku                   0.54   \n",
       "\n",
       "                                                                          lev_dist  \\\n",
       "                                                                              mean   \n",
       "type                                               model                             \n",
       "nosm_expert_recover_instructions_prompt_gpt-4-t... gpt-4-turbo-preview      148.62   \n",
       "full_context_prompt_gpt-4-turbo-preview            gpt-4-turbo-preview      151.38   \n",
       "expert_recover_pub_instructions_prompt_gpt-4-tu... gpt-4-turbo-preview      153.43   \n",
       "expert_recover_instructions_prompt_claude-3-opu... claude-3-opus-20240229   183.24   \n",
       "nosm_full_context_prompt_gpt-4-turbo-preview       gpt-4-turbo-preview      160.43   \n",
       "...                                                                            ...   \n",
       "nosm_expert_recover_publication_prompt_haiku       haiku                    438.71   \n",
       "expert_recover_publication_text_prompt_haiku       haiku                    581.00   \n",
       "expert_recover_text_prompt_haiku                   haiku                    785.76   \n",
       "nosm_expert_recover_text_prompt_haiku              haiku                    630.95   \n",
       "nosm_expert_recover_publication_text_prompt_haiku  haiku                    605.81   \n",
       "\n",
       "                                                                                  \n",
       "                                                                             50%  \n",
       "type                                               model                          \n",
       "nosm_expert_recover_instructions_prompt_gpt-4-t... gpt-4-turbo-preview      77.0  \n",
       "full_context_prompt_gpt-4-turbo-preview            gpt-4-turbo-preview      80.0  \n",
       "expert_recover_pub_instructions_prompt_gpt-4-tu... gpt-4-turbo-preview      80.0  \n",
       "expert_recover_instructions_prompt_claude-3-opu... claude-3-opus-20240229   84.0  \n",
       "nosm_full_context_prompt_gpt-4-turbo-preview       gpt-4-turbo-preview      87.0  \n",
       "...                                                                          ...  \n",
       "nosm_expert_recover_publication_prompt_haiku       haiku                   396.0  \n",
       "expert_recover_publication_text_prompt_haiku       haiku                   526.0  \n",
       "expert_recover_text_prompt_haiku                   haiku                   543.0  \n",
       "nosm_expert_recover_text_prompt_haiku              haiku                   549.0  \n",
       "nosm_expert_recover_publication_text_prompt_haiku  haiku                   575.0  \n",
       "\n",
       "[72 rows x 6 columns]"
      ]
     },
     "execution_count": 17,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "performance_eval2 = performance_eval.copy()\n",
    "performance_eval2['type'] = performance_eval2['type'].str.replace(\"claude-3-haiku-20240307\", \"haiku\").replace(\"gpt-3.5-turbo\", \"gpt-3.5\")\n",
    "performance_eval2['model'] = performance_eval2['type'].str.split('_').str[-1]\n",
    "#performance_eval2 = performance_eval2.loc[performance_eval2['model'].str.contains('gpt-4')]\n",
    "performance_eval2.drop(columns = 'File Name').groupby(['type', 'model']).describe().filter(regex = '50|mean').round(2).sort_values(('lev_dist', '50%'))"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 24,
   "metadata": {},
   "outputs": [
    {
     "data": {
      "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></th>\n",
       "      <th>CER</th>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>model</th>\n",
       "      <th>type</th>\n",
       "      <th></th>\n",
       "    </tr>\n",
       "  </thead>\n",
       "  <tbody>\n",
       "    <tr>\n",
       "      <th rowspan=\"2\" valign=\"top\">claude-3-opus-20240229</th>\n",
       "      <th>nosm_full_context_prompt_claude-3-opus-20240229</th>\n",
       "      <td>0.041680</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>expert_recover_instructions_prompt_claude-3-opus-20240229</th>\n",
       "      <td>0.048680</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th rowspan=\"2\" valign=\"top\">gpt-3.5-turbo</th>\n",
       "      <th>expert_basic_prompt_gpt-3.5-turbo</th>\n",
       "      <td>0.052180</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>expert_recover_publication_prompt_gpt-3.5-turbo</th>\n",
       "      <td>0.054725</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th rowspan=\"2\" valign=\"top\">gpt-4-turbo-preview</th>\n",
       "      <th>expert_recover_pub_instructions_prompt_gpt-4-turbo-preview</th>\n",
       "      <td>0.045180</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>expert_recover_instructions_prompt_gpt-4-turbo-preview</th>\n",
       "      <td>0.048680</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th rowspan=\"2\" valign=\"top\">haiku</th>\n",
       "      <th>full_context_prompt_haiku</th>\n",
       "      <td>0.093135</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>nosm_expert_basic_prompt_haiku</th>\n",
       "      <td>0.093254</td>\n",
       "    </tr>\n",
       "  </tbody>\n",
       "</table>\n",
       "</div>"
      ],
      "text/plain": [
       "                                                                                CER\n",
       "model                  type                                                        \n",
       "claude-3-opus-20240229 nosm_full_context_prompt_claude-3-opus-20240229     0.041680\n",
       "                       expert_recover_instructions_prompt_claude-3-opu...  0.048680\n",
       "gpt-3.5-turbo          expert_basic_prompt_gpt-3.5-turbo                   0.052180\n",
       "                       expert_recover_publication_prompt_gpt-3.5-turbo     0.054725\n",
       "gpt-4-turbo-preview    expert_recover_pub_instructions_prompt_gpt-4-tu...  0.045180\n",
       "                       expert_recover_instructions_prompt_gpt-4-turbo-...  0.048680\n",
       "haiku                  full_context_prompt_haiku                           0.093135\n",
       "                       nosm_expert_basic_prompt_haiku                      0.093254"
      ]
     },
     "execution_count": 24,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "result = performance_eval2.drop(columns='File Name') \\\n",
    "    .groupby(['model', 'type'])['CER'] \\\n",
    "    .median() \\\n",
    "    .reset_index() \\\n",
    "    .sort_values('CER') \\\n",
    "    .groupby('model') \\\n",
    "    .head(2) \\\n",
    "    .set_index(['model', 'type']).sort_values('model')\n",
    "\n",
    "\n",
    "#result['prompt'] = result['type'].str.split('_').str[0]\n",
    "result"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 48,
   "metadata": {},
   "outputs": [],
   "source": [
    "import matplotlib.ticker as ticker"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 50,
   "metadata": {},
   "outputs": [
    {
     "name": "stderr",
     "output_type": "stream",
     "text": [
      "/tmp/ipykernel_70231/532594401.py:19: UserWarning: set_ticklabels() should only be used with a fixed number of ticks, i.e. after set_ticks() or using a FixedLocator.\n",
      "  ax.set_xticklabels(ax.get_xticklabels(), rotation=45, ha='right')\n",
      "/tmp/ipykernel_70231/532594401.py:19: UserWarning: set_ticklabels() should only be used with a fixed number of ticks, i.e. after set_ticks() or using a FixedLocator.\n",
      "  ax.set_xticklabels(ax.get_xticklabels(), rotation=45, ha='right')\n"
     ]
    },
    {
     "data": {
      "image/png": "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",
      "text/plain": [
       "<Figure size 1127.74x1000 with 4 Axes>"
      ]
     },
     "metadata": {},
     "output_type": "display_data"
    }
   ],
   "source": [
    "result = performance_eval2.drop(columns='File Name') \\\n",
    "    .groupby(['model', 'type'])['CER'] \\\n",
    "    .median() \\\n",
    "    .reset_index() \n",
    "\n",
    "\n",
    "result['prompt'] = result['type'].str.rsplit('_', n=2).str[0].str.replace(\"nosm_\", \"\")\n",
    "result['system message'] = np.where(result['type'].str.contains('nosm'), True, False)\n",
    "\n",
    "result = result.sort_values(['prompt','system message'])\n",
    "\n",
    "# Define the order in which 'prompt' categories should appear\n",
    "#order = result['prompt'].unique()\n",
    "\n",
    "\n",
    "g = sns.catplot(data=result, x='prompt', y='CER', hue='system message', kind='bar',  col='model', col_wrap= 2,  sharey=False)\n",
    "\n",
    "for ax in g.axes.flat:\n",
    "    ax.set_xticklabels(ax.get_xticklabels(), rotation=45, ha='right')\n",
    "\n",
    "plt.show()\n"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 55,
   "metadata": {},
   "outputs": [],
   "source": [
    "save_appendix"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 64,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "claude-3-opus-20240229\n",
      "gpt-3.5-turbo\n",
      "gpt-4-turbo-preview\n",
      "haiku\n"
     ]
    }
   ],
   "source": [
    "# Loop through each unique model and create a separate plot\n",
    "for model in result['model'].unique():\n",
    "    # Filter data for the current model\n",
    "    print(model)\n",
    "    model_data = result[result['model'] == model]\n",
    "    \n",
    "    # Create a bar plot\n",
    "    plt.figure(figsize=(10, 6))\n",
    "    sns.barplot(data=model_data, x='prompt', y='CER', hue='system message')\n",
    "    \n",
    "    # Set the title and labels\n",
    "    plt.title(f'CER by Prompt Type for Model: {model}')\n",
    "    plt.xlabel('Prompt')\n",
    "    plt.ylabel('CER')\n",
    "    plt.xticks(rotation=25, ha='right')\n",
    "    plt.tight_layout()\n",
    "    \n",
    "    # Save the plot\n",
    "    fig_filename = os.path.join(save_appendix, f\"cer_{model.lower()}.png\")\n",
    "    plt.savefig(fig_filename)\n",
    "    plt.close()  # Close the figure to free up memory"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## Conclusions of the prompt test\n",
    "\n",
    "It appears that the placing the prompt after the text instead of using the system prompt gives the best results. However, the prompts did give significantlty different performance.  I think that perhaps using the `full_context_prompt` and the `expert_recover_publication_prompt` with no system message and the prompt after the text may be the best option. This will require twice as much compute as I was planning to use."
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "# Evaluating models on the test set\n",
    "\n",
    "Having identified two different prompts and that the prompts appear to work better when places after the text we can now compare the different models\n",
    "\n",
    "\n",
    "The below code creates the basic configuration dictionaries for each model and then fills in the with the two different prompt messages creating a single list of all basic prompt/model configurations. It then calls all the LLM's and saves the results.\n",
    "This works in series so takes a while."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 65,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Observed median difference: 1.7362332392678645\n",
      "P-Value from Bootstrap Test: 0.5326\n"
     ]
    }
   ],
   "source": [
    "#There is no significant difference between nosm and sm so, who knows.\n",
    "\n",
    "test2 = test.groupby('type')['lev_dist'].median().reset_index().copy()\n",
    "\n",
    "test2['system_message'] = ~test2['type'].str.startswith('nosm')\n",
    "\n",
    "test2['new_type'] = test2['type'].str.replace('nosm_', '', regex=True)\n",
    "\n",
    "# Splitting data based on 'system_message'\n",
    "data_true = test2[test2['system_message'] == True]\n",
    "data_false = test2[test2['system_message'] == False]\n",
    "\n",
    "# Merging on 'new_type'\n",
    "merged_data = pd.merge(data_true, data_false, on='new_type', suffixes=('_true', '_false'))\n",
    "\n",
    "# Bootstrap parameters\n",
    "n_bootstrap = 10000\n",
    "differences = []\n",
    "\n",
    "for _ in range(n_bootstrap):\n",
    "    # Resample the merged data with replacement\n",
    "    sample = merged_data.sample(n=len(merged_data), replace=True)\n",
    "    # Compute differences\n",
    "    diff = sample['lev_dist_true'] - sample['lev_dist_false']\n",
    "    # Calculate the median of these differences\n",
    "    differences.append(diff.median())\n",
    "\n",
    "# Calculate observed difference in medians from the original data\n",
    "obs_diff = (merged_data['lev_dist_true'] - merged_data['lev_dist_false']).median()\n",
    "\n",
    "# Estimate the p-value\n",
    "p_value = np.mean(np.abs(differences) >= np.abs(obs_diff))\n",
    "\n",
    "print(\"Observed median difference:\", obs_diff)\n",
    "print(\"P-Value from Bootstrap Test:\", p_value)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": []
  }
 ],
 "metadata": {
  "kernelspec": {
   "display_name": ".venv",
   "language": "python",
   "name": "python3"
  },
  "language_info": {
   "codemirror_mode": {
    "name": "ipython",
    "version": 3
   },
   "file_extension": ".py",
   "mimetype": "text/x-python",
   "name": "python",
   "nbconvert_exporter": "python",
   "pygments_lexer": "ipython3",
   "version": "3.11.6"
  }
 },
 "nbformat": 4,
 "nbformat_minor": 2
}