{
 "cells": [
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "# Sid-GPT demo\n",
    "A 92M-parameter GPT built from scratch that rewrites weak PM resume bullets into quantified, action-led ones.\n",
    "\n",
    "Run the cells top to bottom (Runtime \u2192 Run all). The last cell gives you a public link to a small web app."
   ]
  },
  {
   "cell_type": "code",
   "metadata": {},
   "execution_count": null,
   "outputs": [],
   "source": [
    "!pip -q install torch tokenizers huggingface_hub gradio"
   ]
  },
  {
   "cell_type": "code",
   "metadata": {},
   "execution_count": null,
   "outputs": [],
   "source": [
    "import sys, torch\n",
    "from huggingface_hub import snapshot_download\n",
    "path = snapshot_download(\"Sidhartha-Rajput/sid-gpt-resume-llm\")\n",
    "sys.path.insert(0, path)\n",
    "from model import GPT\n",
    "from tokenizers import ByteLevelBPETokenizer\n",
    "\n",
    "tok = ByteLevelBPETokenizer(f\"{path}/tokenizer/vocab.json\", f\"{path}/tokenizer/merges.txt\")\n",
    "ckpt = torch.load(f\"{path}/sft.pt\", map_location=\"cpu\", weights_only=False)\n",
    "model = GPT(ckpt[\"config\"]); model.load_state_dict(ckpt[\"model\"]); model.eval()\n",
    "UID, AID, EOT = (tok.token_to_id(t) for t in (\"<|user|>\", \"<|assistant|>\", \"<|endoftext|>\"))\n",
    "TEMPLATE = \"Turn this into a strong, quantified PM resume bullet:\\n{}\"\n",
    "\n",
    "def rewrite(bullet, temperature=0.5, top_k=20, max_new_tokens=70):\n",
    "    ids = [UID] + tok.encode(TEMPLATE.format(bullet.strip())).ids + [AID]\n",
    "    with torch.no_grad():\n",
    "        out = model.generate(torch.tensor([ids]), max_new_tokens=int(max_new_tokens),\n",
    "                             temperature=float(temperature), top_k=int(top_k))\n",
    "    new = out[0, len(ids):].tolist()\n",
    "    if EOT in new: new = new[:new.index(EOT)]\n",
    "    return tok.decode(new).strip()\n",
    "\n",
    "print(rewrite(\"Responsible for managing the product roadmap for the mobile app.\"))"
   ]
  },
  {
   "cell_type": "code",
   "metadata": {},
   "execution_count": null,
   "outputs": [],
   "source": [
    "import gradio as gr\n",
    "gr.Interface(\n",
    "    fn=rewrite,\n",
    "    inputs=[gr.Textbox(label=\"Your resume bullet\", lines=3),\n",
    "            gr.Slider(0.2, 1.2, value=0.5, label=\"Temperature\"),\n",
    "            gr.Slider(1, 100, value=20, step=1, label=\"Top-k\"),\n",
    "            gr.Slider(20, 120, value=70, step=5, label=\"Max new tokens\")],\n",
    "    outputs=gr.Textbox(label=\"Sid-GPT rewrite\", lines=4),\n",
    "    title=\"Sid-GPT: PM resume bullet rewriter\",\n",
    ").launch(share=True)"
   ]
  }
 ],
 "metadata": {
  "kernelspec": {
   "display_name": "Python 3",
   "name": "python3"
  },
  "language_info": {
   "name": "python"
  }
 },
 "nbformat": 4,
 "nbformat_minor": 5
}