{
 "cells": [
  {
   "cell_type": "code",
   "execution_count": 1,
   "metadata": {},
   "outputs": [],
   "source": [
    "import pandas as pd\n",
    "import os\n",
    "import json\n",
    "import datetime\n",
    "\n",
    "time_now = datetime.datetime.now().strftime(\"%Y-%m-%dT%H-%M-%S\")"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 2,
   "metadata": {},
   "outputs": [],
   "source": [
    "df = pd.read_csv(\"results.csv\")\n",
    "new_df = df.groupby([\"model\", \"problem\"], as_index=False)[['weighted_accuracy']].sum()"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 3,
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/plain": [
       "array(['Claude 2', 'Claude Instant', 'GPT 3.5 Turbo', 'GPT 4 Turbo',\n",
       "       'MPT-30b', 'Mistral-7b', 'PaLM 2', 'Phi-1.5', 'Phi-2', 'Qwen-14b',\n",
       "       'Vicuna-13b', 'Yi-34b'], dtype=object)"
      ]
     },
     "execution_count": 3,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "new_df.model.unique()"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "- https://grabon.com/blog/claude-users-statistics/\n",
    "- https://medium.com/@seanbetts/peering-inside-gpt-4-understanding-its-mixture-of-experts-moe-architecture-2a42eb8bdcb3\n",
    "- https://www.cnbc.com/2023/05/16/googles-palm-2-uses-nearly-five-times-more-text-data-than-predecessor.html"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 4,
   "metadata": {},
   "outputs": [],
   "source": [
    "open_models = {\n",
    "    \"Yi-34b\": \"01-ai/Yi-34B-Chat\",\n",
    "    \"Mistral-7b\": \"mistralai/Mistral-7B-Instruct-v0.1\",\n",
    "    \"Vicuna-13b\": \"lmsys/vicuna-13b-v1.3\",\n",
    "    \"Phi-1.5\": \"microsoft/phi-1_5\",\n",
    "    \"MPT-30b\": \"mosaicml/mpt-30b-instruct\",\n",
    "    \"Phi-2\": \"microsoft/phi-2\",\n",
    "    \"Qwen-14b\": \"Qwen/Qwen-14B-Chat\",\n",
    "}\n",
    "\n",
    "model_params = {\n",
    "    'Yi-34b': ('torch.bfloat16', 34.389),\n",
    "    'Mistral-7b': ('torch.bfloat16', 7.242),\n",
    "    'Vicuna-13b': ('torch.float16', 13.0),\n",
    "    'Phi-1.5': ('torch.float16', 1.3),\n",
    "    'MPT-30b': ('torch.bfloat16', 30.0),\n",
    "    'Phi-2': ('torch.float16', 2.78),\n",
    "    'Qwen-14b': ('torch.bfloat16', 14.167),\n",
    "    'Claude 2': ('?', 176),\n",
    "    'Claude Instant': ('?', 60),\n",
    "    \"GPT 3.5 Turbo\": ('?', 175),\n",
    "    \"GPT 4 Turbo\": ('?', 1760),\n",
    "    'PaLM 2': ('?', 340),\n",
    "}"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 5,
   "metadata": {},
   "outputs": [],
   "source": [
    "def result_export(model_df, model_name):\n",
    "    model_df = model_df.set_index(\"problem\")\n",
    "    model_df = model_df.drop(columns=[\"model\"])\n",
    "    model_df = model_df.to_dict(orient=\"index\")\n",
    "    convert_problem_name = lambda x: x.replace(\"_Results\", \"\").replace(\"Results\", \"\").replace(\"bsp\", \"sas\").upper()\n",
    "    model_df = {convert_problem_name(k): v for k, v in model_df.items()}\n",
    "    return model_df"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 6,
   "metadata": {},
   "outputs": [],
   "source": [
    "for model in new_df.model.unique():   \n",
    "    model_dir = open_models[model] if model in open_models else model.replace(\" \", \"-\")\n",
    "    # os.system(f\"rm -rf {model_dir.split('/')[0]}\")\n",
    "    os.makedirs(f\"{model_dir}\", exist_ok=True)\n",
    "    model_df = new_df[new_df[\"model\"] == model]\n",
    "    model_result = result_export(model_df, model)\n",
    "    model_result = {\n",
    "        \"config\": {\n",
    "            \"model_name\": model_dir, \n",
    "            \"model_type\": \"open-source\" if model in open_models else \"close-source\",\n",
    "            \"model_dtype\": model_params[model][0] if model in model_params else \"?\",\n",
    "            \"num_params\": model_params[model][1] if model in model_params else 0,\n",
    "        },\n",
    "        \"results\": model_result\n",
    "    }\n",
    "    with open(f\"{model_dir}/results_{time_now}.json\", \"w\") as f:\n",
    "        json.dump(model_result, f, indent=4)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": []
  }
 ],
 "metadata": {
  "kernelspec": {
   "display_name": "llm_reason",
   "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": "3.10.13"
  }
 },
 "nbformat": 4,
 "nbformat_minor": 2
}