{
 "cells": [
  {
   "cell_type": "code",
   "execution_count": 1,
   "id": "73e2d6cb-82b5-42cc-919d-dceaaf1f09f7",
   "metadata": {},
   "outputs": [
    {
     "name": "stderr",
     "output_type": "stream",
     "text": [
      "/home/apapagiannis/update_2025/eurovoc_training/.venv/lib/python3.10/site-packages/tqdm/auto.py:21: TqdmWarning: IProgress not found. Please update jupyter and ipywidgets. See https://ipywidgets.readthedocs.io/en/stable/user_install.html\n",
      "  from .autonotebook import tqdm as notebook_tqdm\n",
      "Loading weights: 100%|██████████| 197/197 [00:00<00:00, 8783.65it/s]\n",
      "[transformers] \u001b[1mBertModel LOAD REPORT\u001b[0m from: ./merged_model\n",
      "Key                                        | Status     | \n",
      "-------------------------------------------+------------+-\n",
      "cls.predictions.transform.dense.bias       | UNEXPECTED | \n",
      "cls.predictions.transform.LayerNorm.bias   | UNEXPECTED | \n",
      "cls.predictions.transform.LayerNorm.weight | UNEXPECTED | \n",
      "cls.predictions.transform.dense.weight     | UNEXPECTED | \n",
      "cls.predictions.bias                       | UNEXPECTED | \n",
      "pooler.dense.bias                          | MISSING    | \n",
      "pooler.dense.weight                        | MISSING    | \n",
      "\n",
      "Notes:\n",
      "- UNEXPECTED:\tcan be ignored when loading from different task/architecture; not ok if you expect identical arch.\n",
      "- MISSING:\tthose params were newly initialized because missing from the checkpoint. Consider training on your downstream task.\n"
     ]
    },
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "energy vs energy:   0.7808\n",
      "energy vs criminal: 0.5920\n"
     ]
    }
   ],
   "source": [
    "import torch\n",
    "from transformers import AutoModel, AutoTokenizer\n",
    "\n",
    "model_dir = \"./merged_model\"\n",
    "\n",
    "# What you actually want — embedding quality\n",
    "model = AutoModel.from_pretrained(model_dir)  # NOT AutoModelForMaskedLM\n",
    "tokenizer = AutoTokenizer.from_pretrained(model_dir)\n",
    "\n",
    "def embed(text):\n",
    "    inputs = tokenizer(text, return_tensors=\"pt\", truncation=True, max_length=512)\n",
    "    with torch.no_grad():\n",
    "        out = model(**inputs)\n",
    "    return out.last_hidden_state[:, 0, :]  # CLS token\n",
    "\n",
    "# Test: similar legal concepts should be closer than unrelated ones\n",
    "from torch.nn.functional import cosine_similarity\n",
    "\n",
    "e1 = embed(\"renewable energy regulation\")\n",
    "e2 = embed(\"solar power and wind policy\")  # should be close\n",
    "e3 = embed(\"criminal sentencing guidelines\")       # should be far\n",
    "\n",
    "print(f\"energy vs energy:   {cosine_similarity(e1, e2).item():.4f}\")  # expect ~high\n",
    "print(f\"energy vs criminal: {cosine_similarity(e1, e3).item():.4f}\")  # expect ~lower"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "id": "94556c9a-bd65-489f-b2ab-d44b9b0812c7",
   "metadata": {},
   "outputs": [],
   "source": []
  }
 ],
 "metadata": {
  "kernelspec": {
   "display_name": "eurovoc_training",
   "language": "python",
   "name": "my-venv"
  },
  "language_info": {
   "codemirror_mode": {
    "name": "ipython",
    "version": 3
   },
   "file_extension": ".py",
   "mimetype": "text/x-python",
   "name": "python",
   "nbconvert_exporter": "python",
   "pygments_lexer": "ipython3",
   "version": "3.10.12"
  }
 },
 "nbformat": 4,
 "nbformat_minor": 5
}