Duplicate from Ory999/Assignment_2

.gitattributes

@@ -0,0 +1,35 @@
+*.7z filter=lfs diff=lfs merge=lfs -text
+*.arrow filter=lfs diff=lfs merge=lfs -text
+*.bin filter=lfs diff=lfs merge=lfs -text
+*.bz2 filter=lfs diff=lfs merge=lfs -text
+*.ckpt filter=lfs diff=lfs merge=lfs -text
+*.ftz filter=lfs diff=lfs merge=lfs -text
+*.gz filter=lfs diff=lfs merge=lfs -text
+*.h5 filter=lfs diff=lfs merge=lfs -text
+*.joblib filter=lfs diff=lfs merge=lfs -text
+*.lfs.* filter=lfs diff=lfs merge=lfs -text
+*.mlmodel filter=lfs diff=lfs merge=lfs -text
+*.model filter=lfs diff=lfs merge=lfs -text
+*.msgpack filter=lfs diff=lfs merge=lfs -text
+*.npy filter=lfs diff=lfs merge=lfs -text
+*.npz filter=lfs diff=lfs merge=lfs -text
+*.onnx filter=lfs diff=lfs merge=lfs -text
+*.ot filter=lfs diff=lfs merge=lfs -text
+*.parquet filter=lfs diff=lfs merge=lfs -text
+*.pb filter=lfs diff=lfs merge=lfs -text
+*.pickle filter=lfs diff=lfs merge=lfs -text
+*.pkl filter=lfs diff=lfs merge=lfs -text
+*.pt filter=lfs diff=lfs merge=lfs -text
+*.pth filter=lfs diff=lfs merge=lfs -text
+*.rar filter=lfs diff=lfs merge=lfs -text
+*.safetensors filter=lfs diff=lfs merge=lfs -text
+saved_model/**/* filter=lfs diff=lfs merge=lfs -text
+*.tar.* filter=lfs diff=lfs merge=lfs -text
+*.tar filter=lfs diff=lfs merge=lfs -text
+*.tflite filter=lfs diff=lfs merge=lfs -text
+*.tgz filter=lfs diff=lfs merge=lfs -text
+*.wasm filter=lfs diff=lfs merge=lfs -text
+*.xz filter=lfs diff=lfs merge=lfs -text
+*.zip filter=lfs diff=lfs merge=lfs -text
+*.zst filter=lfs diff=lfs merge=lfs -text
+*tfevents* filter=lfs diff=lfs merge=lfs -text

README.md

@@ -0,0 +1,68 @@
+# Active Learning for Patent Classification: Green Technology Identification
+
+This repository contains an Active Learning pipeline designed to classify patent claims into "Green" (Climate Change Mitigation) and "Not Green" categories based on the CPC Y02 taxonomy.
+
+## Project Overview
+This project implements a Human-in-the-Loop (HITL) workflow to refine a patent classifier. By combining automated silver labeling with expert human validation of high-uncertainty samples, we improve the model's ability to identify complex environmental technologies.
+
+## Part A: Baseline Model Evaluation
+The baseline model utilized **PatentSBERTa** for feature extraction and a **Logistic Regression** classifier trained on 2,000 silver-labeled samples.
+
+**Baseline Performance Metrics:**
+* **Macro F1-Score:** 0.7488
+* **Macro Precision:** 0.7489
+* **Macro Recall:** 0.7488
+
+The baseline demonstrated reasonable accuracy but lacked nuance in identifying technologies outside of primary renewables like wind and solar.
+
+## Part B: Uncertainty Sampling
+Using Uncertainty Sampling, 100 high-risk claims (where predicted probability $p \approx 0.5$) were extracted for human review. The formula used for uncertainty was $u = 1 - 2 \cdot |p - 0.5|$, where a score of 1.0 represents maximum uncertainty.
+
+> **Note on Part B:** Due to a kernel restart during the workflow, the script generates a reproduction file named `hitl_green_100_REPRODUCED.csv`. This ensures that the original labeling progress is protected while still providing the required metrics for the uncertainty scores (Min: 0.9959, Max: 1.0000).
+
+## Part C: Human-in-the-Loop (HITL) Reporting
+An LLM provided initial suggestions and rationales, which were then validated or overridden by a human expert. The LLM used was GPT-OSS-20B which was locally run via LM studio on a 9070 XT 16Gb Vram AMD card the Vulkan software. Temperature set to 0.1.
+
+**Labeling Statistics:**
+* **Total Claims Labeled:** 100
+* **Human Overrides:** 3 - detected mistakes where LLM failed to classify a technology that was green.
+* **Agreement Rate:** 97.0%
+
+### Examples of Human Overrides
+A critical component of this project was correcting model biases. Below are representative examples where the human expert overrode the LLM suggestion:
+
+1. **Nuclear Reactor Components (Doc 9064607)**
+* **LLM Suggestion:** 0
+* **Human Final:** 1
+* **Rationale:** Nuclear Energy is recognized as a low-carbon energy source for climate change mitigation under CPC Y02E 30/00. The LLM displayed a bias toward renewable-only definitions.
+
+2. **Wastewater Treatment Flocculants (Doc 9328007)**
+* **LLM Suggestion:** 0
+* **Human Final:** 1
+* **Rationale:** Technologies reducing environmental impact in waste management are categorized under Y02W. The human expert corrected the LLM's exclusion of wastewater treatment.
+
+3. **Pollution Abatement Reactor (Doc 8833279)**
+* **LLM Suggestion:** 0
+* **Human Final:** 1
+* **Rationale:** Technologies for atmospheric pollution abatement (Y02A) are included in the green patent taxonomy. The human expert recognized this broader scope.
+
+## Part D: Active Learning Evaluation
+The model was retrained using the original training set plus the 100 human-labeled "Gold" samples.
+
+**Final Performance Comparison:**
+
+| Metric | Baseline (Part A) | Active (Part D) | Improvement |
+| :--- | :--- | :--- | :--- |
+| **Precision** | 0.7489 | 0.7473 | -0.0015 |
+| **Recall** | 0.7488 | 0.7467 | -0.0021 |
+| **F1-score** | 0.7488 | 0.7465 | -0.0023 |
+
+
+
+## Analysis of Results
+The Active Learning loop provided a qualitative improvement by resolving high-uncertainty samples. While the numerical F1-score showed a minor decrease of approximately 0.2%, this is expected when introducing complex, "expert" labels that challenge the simple patterns learned from silver labels. The model gained specific robustness in identifying "hidden" green categories like Nuclear and Wastewater management, effectively trading a small amount of generic accuracy for higher expert precision in contested technology areas.
+## Files in Repository
+
+* **assignment_2.ipynb**: The comprehensive Python implementation including data processing, baseline training, uncertainty sampling, human labeling loops, and the final active learning evaluation.
+* **hitl_green_100.csv**: The 100-sample "Gold Label" dataset generated during the human-in-the-loop phase, featuring model predictions, LLM rationales, and final human-verified labels.
+* **patents_50k_green.parquet**: The balanced silver-standard training dataset consisting of 50,000 patent claims, which is required for the baseline and active learning model training.

assignment_2.ipynb

@@ -0,0 +1,1319 @@
+{
+ "cells": [
+  {
+   "cell_type": "markdown",
+   "id": "4312be94",
+   "metadata": {},
+   "source": [
+    "# Part A"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": 26,
+   "id": "499c83de",
+   "metadata": {},
+   "outputs": [
+    {
+     "name": "stdout",
+     "output_type": "stream",
+     "text": [
+      "Data already exists in 'patents_data_raw'. Skipping download.\n"
+     ]
+    }
+   ],
+   "source": [
+    "import os\n",
+    "from huggingface_hub import snapshot_download\n",
+    "\n",
+    "# Check if the folder already exists and is not empty\n",
+    "folder_name = \"patents_data_raw\"\n",
+    "\n",
+    "# Check if folder exists and if it has any files inside\n",
+    "if os.path.exists(folder_name) and any(os.scandir(folder_name)):\n",
+    "    print(f\"Data already exists in '{folder_name}'. Skipping download.\")\n",
+    "    local_folder = os.path.abspath(folder_name)\n",
+    "else:\n",
+    "    #Download only if missing\n",
+    "    print(f\"Downloading dataset files to '{folder_name}'\")\n",
+    "    try:\n",
+    "        local_folder = snapshot_download(\n",
+    "            repo_id=\"AI-Growth-Lab/patents_claims_1.5m_traim_test\", \n",
+    "            repo_type=\"dataset\",\n",
+    "            local_dir=folder_name,\n",
+    "            ignore_patterns=[\"*.git*\"]\n",
+    "        )\n",
+    "        print(f\"Success! Files downloaded to: {local_folder}\")\n",
+    "    except Exception as e:\n",
+    "        print(f\"Download failed: {e}\")"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": 27,
+   "id": "2a1e5f1b",
+   "metadata": {},
+   "outputs": [
+    {
+     "name": "stdout",
+     "output_type": "stream",
+     "text": [
+      "Found existing processed file: patents_50k_green.parquet\n",
+      "   Skipping filtering and merging to save time.\n"
+     ]
+    }
+   ],
+   "source": [
+    "from datasets import load_dataset, concatenate_datasets, disable_progress_bar\n",
+    "import pandas as pd\n",
+    "import datasets\n",
+    "\n",
+    "# Silience Hugging Face logs for cleaner output\n",
+    "disable_progress_bar()\n",
+    "datasets.utils.logging.set_verbosity_error()\n",
+    "\n",
+    "output_filename = \"patents_50k_green.parquet\"\n",
+    "\n",
+    "# Check if this work is already done\n",
+    "if os.path.exists(output_filename):\n",
+    "    print(f\"Found existing processed file: {output_filename}\")\n",
+    "    print(f\"   Skipping filtering and merging to save time.\")\n",
+    "else:\n",
+    "    print(\"1. Loading dataset from local cache...\")\n",
+    "    # Point to the local folder\n",
+    "    dataset_full = load_dataset(\"./patents_data_raw\", split=\"train\")\n",
+    "\n",
+    "    print(f\"   Dataset loaded. Total rows: {len(dataset_full):,}\")\n",
+    "\n",
+    "    # Identify Green Columns\n",
+    "    all_cols = dataset_full.column_names\n",
+    "    y02_cols = [c for c in all_cols if c.startswith(\"Y02\")]\n",
+    "    print(f\"   Found {len(y02_cols)} Green Patent (Y02) indicator columns.\")\n",
+    "\n",
+    "    # Filtering Logic\n",
+    "    print(\"2. Filtering for 25,000 Green patents...\")\n",
+    "    dataset_green = dataset_full.filter(\n",
+    "        lambda x: any(x[col] == 1 for col in y02_cols),\n",
+    "        num_proc=1\n",
+    "    ).shuffle(seed=42).select(range(25000))\n",
+    "\n",
+    "    print(\"3. Filtering for 25,000 Non-Green patents...\")\n",
+    "    dataset_not_green = dataset_full.filter(\n",
+    "        lambda x: all(x[col] == 0 for col in y02_cols),\n",
+    "        num_proc=1\n",
+    "    ).shuffle(seed=42).select(range(25000))\n",
+    "\n",
+    "    # 4. Add \"is_green_silver\" Labels\n",
+    "    print(\"4. Adding silver labels (0/1)...\")\n",
+    "    dataset_green = dataset_green.map(lambda x: {\"is_green_silver\": 1})\n",
+    "    dataset_not_green = dataset_not_green.map(lambda x: {\"is_green_silver\": 0})\n",
+    "\n",
+    "    # 5. Combine and Save\n",
+    "    print(\"5. Merging and saving final Parquet...\")\n",
+    "    final_dataset = concatenate_datasets([dataset_green, dataset_not_green]).shuffle(seed=42)\n",
+    "    final_dataset.to_parquet(output_filename)\n",
+    "\n",
+    "    print(f\"Success! File saved: {output_filename}\")\n",
+    "    print(f\"Total Balanced Rows: {len(final_dataset):,}\")"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": 28,
+   "id": "784cf7cb",
+   "metadata": {},
+   "outputs": [
+    {
+     "name": "stdout",
+     "output_type": "stream",
+     "text": [
+      "Loading patents_50k_green.parquet...\n",
+      " Data Setup Complete\n",
+      "   - train_silver:   2000 rows\n",
+      "   - eval_silver:    5000 rows\n",
+      "   - pool_unlabeled: 43000 rows\n"
+     ]
+    }
+   ],
+   "source": [
+    "from sklearn.model_selection import train_test_split\n",
+    "\n",
+    "# 1. Loading the balanced 50k file\n",
+    "print(\"Loading patents_50k_green.parquet...\")\n",
+    "df = pd.read_parquet(\"patents_50k_green.parquet\")\n",
+    "\n",
+    "# Creating the Splits\n",
+    "# - train_silver: Small initial labeled set to train the baseline (e.g., 2,000 - 5,000 rows)\n",
+    "# - eval_silver: Validation set to test performance (e.g., 5,000 rows)\n",
+    "# - pool_unlabeled: The rest, which you will \"mine\" for high-risk examples.\n",
+    "\n",
+    "# Reserve 5,000 for evaluation\n",
+    "df_eval = df.sample(n=5000, random_state=42)\n",
+    "df_remaining = df.drop(df_eval.index)\n",
+    "\n",
+    "# Reserve 2,000 for the initial \"train_silver\"\n",
+    "df_train_silver = df_remaining.sample(n=2000, random_state=42)\n",
+    "df_pool_unlabeled = df_remaining.drop(df_train_silver.index)\n",
+    "\n",
+    "print(\" Data Setup Complete\")\n",
+    "print(f\"   - train_silver:   {len(df_train_silver)} rows\")\n",
+    "print(f\"   - eval_silver:    {len(df_eval)} rows\")\n",
+    "print(f\"   - pool_unlabeled: {len(df_pool_unlabeled)} rows\")"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": 29,
+   "id": "d56051b7",
+   "metadata": {},
+   "outputs": [
+    {
+     "name": "stdout",
+     "output_type": "stream",
+     "text": [
+      "Starting Part A: Baseline Model Training...\n",
+      "   - Loading 50k dataset...\n",
+      "   - Splits created: Train=2000, Eval=5000, Pool=43000\n",
+      "   - Loading PatentSBERTa model...\n",
+      "   - Generating Training embeddings...\n"
+     ]
+    },
+    {
+     "data": {
+      "application/vnd.jupyter.widget-view+json": {
+       "model_id": "2c18d825d370452ca1eadd3788063e0c",
+       "version_major": 2,
+       "version_minor": 0
+      },
+      "text/plain": [
+       "      Encoding:   0%|          | 0/63 [00:00<?, ?it/s]"
+      ]
+     },
+     "metadata": {},
+     "output_type": "display_data"
+    },
+    {
+     "name": "stdout",
+     "output_type": "stream",
+     "text": [
+      "   - Generating Evaluation embeddings...\n"
+     ]
+    },
+    {
+     "data": {
+      "application/vnd.jupyter.widget-view+json": {
+       "model_id": "fd3be50181a348c8b8e8688ae3568157",
+       "version_major": 2,
+       "version_minor": 0
+      },
+      "text/plain": [
+       "      Encoding:   0%|          | 0/157 [00:00<?, ?it/s]"
+      ]
+     },
+     "metadata": {},
+     "output_type": "display_data"
+    },
+    {
+     "name": "stdout",
+     "output_type": "stream",
+     "text": [
+      "   - Training Logistic Regression...\n",
+      "\n",
+      "=============================================\n",
+      "PART A RESULTS: BASELINE MODEL\n",
+      "=============================================\n",
+      "              precision    recall  f1-score   support\n",
+      "\n",
+      "   Not Green       0.74      0.76      0.75      2493\n",
+      "       Green       0.75      0.74      0.75      2507\n",
+      "\n",
+      "    accuracy                           0.75      5000\n",
+      "   macro avg       0.75      0.75      0.75      5000\n",
+      "weighted avg       0.75      0.75      0.75      5000\n",
+      "\n",
+      "---------------------------------------------\n",
+      "Part A Baseline F1-Score: 0.7488\n",
+      "=============================================\n"
+     ]
+    }
+   ],
+   "source": [
+    "import numpy as np\n",
+    "import torch\n",
+    "from transformers import AutoTokenizer, AutoModel, logging\n",
+    "from sklearn.linear_model import LogisticRegression\n",
+    "from sklearn.metrics import classification_report\n",
+    "from tqdm.auto import tqdm\n",
+    "\n",
+    "# Silence logs\n",
+    "logging.set_verbosity_error()\n",
+    "\n",
+    "print(\"Starting Part A: Baseline Model Training...\")\n",
+    "\n",
+    "# Check for the source file\n",
+    "parquet_file = \"patents_50k_green.parquet\"\n",
+    "if not os.path.exists(parquet_file):\n",
+    "    print(f\"Error: {parquet_file} not found. Please run the Filtering script first.\")\n",
+    "else:\n",
+    "    # Load Data & Create Splits\n",
+    "    print(\"   - Loading 50k dataset...\")\n",
+    "    df = pd.read_parquet(parquet_file)\n",
+    "\n",
+    "    # Creating the Splits (train_silver, eval_silver, pool_unlabeled)\n",
+    "    df_eval = df.sample(n=5000, random_state=42)\n",
+    "    df_remaining = df.drop(df_eval.index)\n",
+    "    df_train = df_remaining.sample(n=2000, random_state=42)\n",
+    "    df_pool = df_remaining.drop(df_train.index)\n",
+    "\n",
+    "    print(f\"   - Splits created: Train={len(df_train)}, Eval={len(df_eval)}, Pool={len(df_pool)}\")\n",
+    "\n",
+    "    # Load PatentSBERTa\n",
+    "    print(\"   - Loading PatentSBERTa model...\")\n",
+    "    model_name = \"AI-Growth-Lab/PatentSBERTa\"\n",
+    "    tokenizer = AutoTokenizer.from_pretrained(model_name)\n",
+    "    model = AutoModel.from_pretrained(model_name)\n",
+    "\n",
+    "    device = torch.device(\"cuda\" if torch.cuda.is_available() else \"cpu\")\n",
+    "    model.to(device); model.eval()\n",
+    "\n",
+    "    # Helper function with clean progress tracking\n",
+    "    def get_embeddings(text_list, batch_size=32):\n",
+    "        all_embeddings = []\n",
+    "        # We keep the tqdm bar small and clean\n",
+    "        for i in tqdm(range(0, len(text_list), batch_size), desc=\"      Encoding\", leave=False):\n",
+    "            batch_texts = text_list[i:i+batch_size]\n",
+    "            inputs = tokenizer(batch_texts, padding=True, truncation=True, max_length=128, return_tensors=\"pt\").to(device)\n",
+    "            with torch.no_grad():\n",
+    "                outputs = model(**inputs)\n",
+    "                embeddings = outputs.last_hidden_state[:, 0, :].cpu().numpy()\n",
+    "                all_embeddings.append(embeddings)\n",
+    "        return np.vstack(all_embeddings)\n",
+    "\n",
+    "    # Generate Embeddings\n",
+    "    print(\"   - Generating Training embeddings...\")\n",
+    "    X_train = get_embeddings(df_train['text'].tolist())\n",
+    "    y_train = df_train['is_green_silver'].values\n",
+    "\n",
+    "    print(\"   - Generating Evaluation embeddings...\")\n",
+    "    X_eval = get_embeddings(df_eval['text'].tolist())\n",
+    "    y_eval = df_eval['is_green_silver'].values\n",
+    "\n",
+    "    # Train Baseline Classifier\n",
+    "    print(\"   - Training Logistic Regression...\")\n",
+    "    clf = LogisticRegression(max_iter=1000, random_state=42)\n",
+    "    clf.fit(X_train, y_train)\n",
+    "\n",
+    "    # FINAL REPORT OUTPUT\n",
+    "    print(\"\\n\" + \"=\"*45)\n",
+    "    print(\"PART A RESULTS: BASELINE MODEL\")\n",
+    "    print(\"=\"*45)\n",
+    "    y_pred = clf.predict(X_eval)\n",
+    "    report = classification_report(y_eval, y_pred, target_names=['Not Green', 'Green'])\n",
+    "    print(report)\n",
+    "    \n",
+    "    # Store Macro F1 for par D\n",
+    "    report_dict = classification_report(y_eval, y_pred, output_dict=True)\n",
+    "    macro_f1 = report_dict['macro avg']['f1-score']\n",
+    "    \n",
+    "    print(\"-\" * 45)\n",
+    "    print(f\"Part A Baseline F1-Score: {macro_f1:.4f}\")\n",
+    "    print(\"=\"*45)"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "id": "99b2f0f6",
+   "metadata": {},
+   "source": [
+    "# Part B"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "id": "9a46f788",
+   "metadata": {},
+   "outputs": [
+    {
+     "name": "stdout",
+     "output_type": "stream",
+     "text": [
+      "--- Starting Part B: Safe Reproduction of Outputs ---\n",
+      "Generating baseline training embeddings...\n"
+     ]
+    },
+    {
+     "data": {
+      "application/vnd.jupyter.widget-view+json": {
+       "model_id": "2fc0f419d2aa404ab6e86ca8e20ce7d0",
+       "version_major": 2,
+       "version_minor": 0
+      },
+      "text/plain": [
+       "Encoding:   0%|          | 0/63 [00:00<?, ?it/s]"
+      ]
+     },
+     "metadata": {},
+     "output_type": "display_data"
+    },
+    {
+     "name": "stdout",
+     "output_type": "stream",
+     "text": [
+      "Generating embeddings for 43000 unlabeled examples...\n"
+     ]
+    },
+    {
+     "data": {
+      "application/vnd.jupyter.widget-view+json": {
+       "model_id": "0d80d393db8542b8af12f15c1d92fddf",
+       "version_major": 2,
+       "version_minor": 0
+      },
+      "text/plain": [
+       "Encoding:   0%|          | 0/1344 [00:00<?, ?it/s]"
+      ]
+     },
+     "metadata": {},
+     "output_type": "display_data"
+    },
+    {
+     "name": "stdout",
+     "output_type": "stream",
+     "text": [
+      "Calculating uncertainty scores...\n",
+      "\n",
+      "========================================\n",
+      "Part B Complete! Outputs successfully shown.\n",
+      "File saved to: hitl_green_100_REPRODUCED.csv\n",
+      "   - Min Uncertainty: 0.9959\n",
+      "   - Max Uncertainty: 1.0000\n",
+      "========================================\n"
+     ]
+    }
+   ],
+   "source": [
+    "print(\"--- Starting Part B: Safe Reproduction of Outputs ---\")\n",
+    "\n",
+    "# Re-initialize Data & Model\n",
+    "df = pd.read_parquet(\"patents_50k_green.parquet\")\n",
+    "df_eval = df.sample(n=5000, random_state=42)\n",
+    "df_remaining = df.drop(df_eval.index)\n",
+    "df_train = df_remaining.sample(n=2000, random_state=42)\n",
+    "df_pool = df_remaining.drop(df_train.index) #unlabeled pool\n",
+    "\n",
+    "# Re-initialize PatentSBERTa\n",
+    "model_name = \"AI-Growth-Lab/PatentSBERTa\"\n",
+    "tokenizer = AutoTokenizer.from_pretrained(model_name)\n",
+    "model = AutoModel.from_pretrained(model_name)\n",
+    "device = torch.device(\"cuda\" if torch.cuda.is_available() else \"cpu\")\n",
+    "model.to(device); model.eval()\n",
+    "\n",
+    "def get_embeddings(text_list, batch_size=32):\n",
+    "    all_embeddings = []\n",
+    "    for i in tqdm(range(0, len(text_list), batch_size), desc=\"Encoding\"):\n",
+    "        batch_texts = text_list[i:i+batch_size]\n",
+    "        inputs = tokenizer(batch_texts, padding=True, truncation=True, max_length=128, return_tensors=\"pt\").to(device)\n",
+    "        with torch.no_grad():\n",
+    "            outputs = model(**inputs)\n",
+    "            all_embeddings.append(outputs.last_hidden_state[:, 0, :].cpu().numpy())\n",
+    "    return np.vstack(all_embeddings)\n",
+    "\n",
+    "# Re-train the Baseline Classifier\n",
+    "print(\"Generating baseline training embeddings...\")\n",
+    "X_train = get_embeddings(df_train['text'].tolist())\n",
+    "clf = LogisticRegression(max_iter=1000, random_state=42)\n",
+    "clf.fit(X_train, df_train['is_green_silver'].values)\n",
+    "\n",
+    "# Generate Embeddings for the Unlabeled Pool\n",
+    "print(f\"Generating embeddings for {len(df_pool)} unlabeled examples...\")\n",
+    "X_pool = get_embeddings(df_pool['text'].tolist())\n",
+    "\n",
+    "# Predict Probabilities and Uncertainty\n",
+    "print(\"Calculating uncertainty scores...\")\n",
+    "probs = clf.predict_proba(X_pool)[:, 1]\n",
+    "uncertainty_scores = 1 - 2 * np.abs(probs - 0.5)\n",
+    "\n",
+    "df_pool['p_green'] = probs\n",
+    "df_pool['u'] = uncertainty_scores\n",
+    "\n",
+    "# Select Top 100\n",
+    "df_high_risk = df_pool.sort_values(by='u', ascending=False).head(100)\n",
+    "\n",
+    "# Format for Export (Changed filename to be safe)\n",
+    "if 'id' in df_high_risk.columns:\n",
+    "    df_high_risk = df_high_risk.rename(columns={'id': 'doc_id'})\n",
+    "else:\n",
+    "    df_high_risk['doc_id'] = df_high_risk.index\n",
+    "\n",
+    "for col in ['llm_green_suggested', 'llm_confidence', 'llm_rationale', 'is_green_human', 'notes']:\n",
+    "    df_high_risk[col] = \"\"\n",
+    "\n",
+    "final_columns = ['doc_id', 'text', 'p_green', 'u', 'llm_green_suggested', 'llm_confidence', 'llm_rationale', 'is_green_human', 'notes']\n",
+    "\n",
+    "# Due to keral restart, a second file which is not used in the next steps is created to avoid confusion. The file is named \"hitl_green_100_REPRODUCED.csv\" to indicate that it is a reproduction of the original \"hitl_green_100.csv\" file, but with a different name to prevent any accidental overwriting or confusion with the original file that may have been generated before the kernel restart.\n",
+    "safe_filename = \"hitl_green_100_REPRODUCED.csv\"\n",
+    "df_high_risk[final_columns].to_csv(safe_filename, index=False)\n",
+    "\n",
+    "print(\"\\n\" + \"=\"*40)\n",
+    "print(f\"Part B Complete! Outputs successfully shown.\")\n",
+    "print(f\"File saved to: {safe_filename}\")\n",
+    "print(f\"   - Min Uncertainty: {df_high_risk['u'].min():.4f}\")\n",
+    "print(f\"   - Max Uncertainty: {df_high_risk['u'].max():.4f}\")\n",
+    "print(\"=\"*40)"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "id": "6317d3a4",
+   "metadata": {},
+   "source": [
+    "# Part C"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "id": "8a7f1e12",
+   "metadata": {},
+   "outputs": [
+    {
+     "name": "stdout",
+     "output_type": "stream",
+     "text": [
+      "--- Part C: HITL Labeling ---\n",
+      "Remaining rows: 68\n",
+      "-----------------------------\n",
+      "\n",
+      "[Row 33] (Uncertainty: 0.9984)\n",
+      "CLAIM: 1. A method of increasing light extraction from a light-emitting diode (LED) device comprising; forming a first n-doped layer on a carrier substrate; forming a Si forming a second n-doped layer on the Si forming an active layer configured to emit light on the second n-doped layer; forming a p-doped ...\n",
+      "------------------------------------------------------------\n",
+      "LLM Says: 0 (High) | Rationale: The claim concerns LED light‑emission enhancement, not climate change mitigation.\n",
+      "Saved.\n",
+      "\n",
+      "[Row 34] (Uncertainty: 0.9984)\n",
+      "CLAIM: 1. A method comprising: identifying valuation data comprising a plurality of estimated asset values corresponding to one or more of location information and property type; identifying a group of two or more characteristics, wherein for each respective real estate investment trust of a plurality of r...\n",
+      "------------------------------------------------------------\n",
+      "LLM Says: 0 (High) | Rationale: The claim concerns financial weighting of real estate trusts, not climate mitigation.\n",
+      "Saved.\n",
+      "\n",
+      "[Row 35] (Uncertainty: 0.9983)\n",
+      "CLAIM: 1. A construction machine, comprising: a lower travel body; an upper slewing body mounted on the lower travel body and having an engine compartment; an engine compartment cover which covers the engine compartment of the upper slewing body; an air filter which collects dust included in outside air ta...\n",
+      "------------------------------------------------------------\n",
+      "LLM Says: 0 (High) | Rationale: The claim describes an air filter system for a construction machine, which does not relate to green or climate change mitigation.\n",
+      "Saved.\n",
+      "\n",
+      "[Row 36] (Uncertainty: 0.9983)\n",
+      "CLAIM: 1. A semiconductor device including a plurality of operation circuits executing operation in synchronization with a clock signal comprising: a control unit for outputting first operation control information and second operation control information for controlling operation executed by the plurality ...\n",
+      "------------------------------------------------------------\n",
+      "LLM Says: 0 (High) | Rationale: The claim describes a generic semiconductor device architecture without any reference to environmental impact or climate change mitigation.\n",
+      "Saved.\n",
+      "\n",
+      "[Row 37] (Uncertainty: 0.9983)\n",
+      "CLAIM: 1. A method of forming a layer over a substrate, the method comprising: receiving data identifying a desired thickness of the layer; using a processor to generate instructions for a printing mechanism to deposit droplets of ink onto the substrate according to the data, the ink carrying material to f...\n",
+      "------------------------------------------------------------\n",
+      "LLM Says: 0 (High) | Rationale: The claim describes a printing method for depositing ink layers, unrelated to climate change mitigation.\n",
+      "Saved.\n",
+      "\n",
+      "[Row 38] (Uncertainty: 0.9982)\n",
+      "CLAIM: 1. An insulated-gate bipolar transistor (IGBT) in a semiconductor substrate, said IGBT comprising: a collector at a bottom surface of said semiconductor substrate, a drift region having a first conductivity type situated over said collector, and a base layer having a second conductivity type opposit...\n",
+      "------------------------------------------------------------\n",
+      "LLM Says: 0 (High) | Rationale: The claim describes a semiconductor device structure, not a technology for greenhouse gas mitigation.\n",
+      "Saved.\n",
+      "\n",
+      "[Row 39] (Uncertainty: 0.9981)\n",
+      "CLAIM: 1. A method of joining two or more articles via slender nanomaterials embedded in a joining medium and interlinked together, the method involving: (i) dispersion of nanomaterials comprising at least one of carbon nanotubes and nanofibers within a solvent, with the weight percent of said nanomaterial...\n",
+      "------------------------------------------------------------\n",
+      "LLM Says: 0 (High) | Rationale: The claim describes a nanomaterial-based joining method, which is unrelated to greenhouse‑gas reduction or climate change mitigation.\n",
+      "Saved.\n",
+      "\n",
+      "[Row 40] (Uncertainty: 0.9981)\n",
+      "CLAIM: 1. A compound having a structure represented by a chemical formula described below:...\n",
+      "------------------------------------------------------------\n",
+      "Asking LM Studio...\n",
+      "LLM output invalid JSON: <|channel|>final <|constrain|>JSON<|message|>{\"suggestion\":0,\"confidence\":\"Low\",\"rationale\":\"The claim merely describes a chemical compound without indicating any environmental or climate‑related application.\"}\n",
+      "LLM Failed (Check settings or Label Manually)\n",
+      "Saved.\n",
+      "\n",
+      "[Row 41] (Uncertainty: 0.9981)\n",
+      "CLAIM: 1. A computer-implemented process comprising: executing, by a computer processor, at least two read threads to read a block of data from a database, each of the read threads having a first wait stat and a second wait stat, the read threads configured to compress data using a dynamic compression rati...\n",
+      "------------------------------------------------------------\n",
+      "LLM Says: 0 (High) | Rationale: The claim describes a data compression and backup process, unrelated to climate change mitigation.\n",
+      "Saved.\n",
+      "\n",
+      "[Row 42] (Uncertainty: 0.9980)\n",
+      "CLAIM: 1. A rotor for a Wankel engine comprising: two axially spaced apart end faces having a generally triangular profile with outwardly arched sides and three circumferentially spaced apex portions; a peripheral face extending between the end faces and defining three flanks, each flank extending between ...\n",
+      "------------------------------------------------------------\n",
+      "LLM Says: 0 (High) | Rationale: The claim describes a mechanical component for a Wankel engine and does not relate to greenhouse gas mitigation or climate change technologies.\n",
+      "Saved.\n",
+      "\n",
+      "[Row 43] (Uncertainty: 0.9980)\n",
+      "CLAIM: 1. A housing apparatus, comprising: a housing casing which surrounds a first cavity and which has multiple side surfaces; a volute housing arranged in an interior of the housing casing, said volute housing having a central through opening for accommodating a compressor wheel of a rotor and for suppl...\n",
+      "------------------------------------------------------------\n",
+      "LLM Says: 0 (Medium) | Rationale: The claim describes a housing apparatus for a compressor wheel, which is unrelated to green or climate change mitigation technologies.\n",
+      "Saved.\n",
+      "\n",
+      "[Row 44] (Uncertainty: 0.9979)\n",
+      "CLAIM: 1. A biodegradable container for a semi-solid composition, comprising: a tube portion comprising a first paper that defines first, second, and third plies forming an open end and a closed end, and a lumen containing the semi-solid composition, wherein the tube portion further comprises a continuous ...\n",
+      "------------------------------------------------------------\n",
+      "LLM Says: 1 (High) | Rationale: The claim describes a fully biodegradable container, indicating an environmental benefit aligned with Green/Climate Change mitigation.\n",
+      "Saved.\n",
+      "\n",
+      "[Row 45] (Uncertainty: 0.9979)\n",
+      "CLAIM: 1. An isolated green sulfur bacterium Chlorobaculum limnaeum strain RK-j-1 deposited at National Institute of Technology and Evaluation Patent Microorganisms Depositary (NPMD) as accession number NITE BP-1202....\n",
+      "------------------------------------------------------------\n",
+      "LLM Says: 0 (Medium) | Rationale: The claim merely describes isolation of a microorganism without any stated application to climate change mitigation.\n",
+      "Saved.\n",
+      "\n",
+      "[Row 46] (Uncertainty: 0.9979)\n",
+      "CLAIM: 1. A method for communicating over allocated resources, comprising: receiving a resource allocation comprising a portion of a resource block over a plurality of bundled transmission time intervals, wherein the portion of the resource block comprises a subset of subcarriers in the resource block with...\n",
+      "------------------------------------------------------------\n",
+      "LLM Says: 0 (High) | Rationale: The claim describes a telecommunications resource allocation method, unrelated to Green/Climate Change mitigation.\n",
+      "Saved.\n",
+      "\n",
+      "[Row 47] (Uncertainty: 0.9979)\n",
+      "CLAIM: 1. A switch system, comprising: a plurality of nodes, wherein each node includes a computational processor and an embedded switch; a plurality of links associated with each node, wherein the plurality of links are configured to connect nodes in the plurality of nodes to create a topology of a switch...\n",
+      "------------------------------------------------------------\n",
+      "LLM Says: 0 (High) | Rationale: The claim describes a network switch architecture, unrelated to climate change mitigation.\n",
+      "Saved.\n",
+      "\n",
+      "[Row 48] (Uncertainty: 0.9977)\n",
+      "CLAIM: 1. A seed of soybean cultivar S100323, wherein a representative of sample seed of said cultivar is deposited under ATCC Accession No. PTA-12317....\n",
+      "------------------------------------------------------------\n",
+      "LLM Says: 0 (High) | Rationale: The claim concerns a soybean seed deposit, not climate‑change mitigation.\n",
+      "Saved.\n",
+      "\n",
+      "[Row 49] (Uncertainty: 0.9976)\n",
+      "CLAIM: 1. A method to accelerate particles into a chamber, comprising: distributing a fluidic substance between electrodes configured at a location proximate a chamber, the electrodes comprising a low work function material; generating a current of ionized particles by applying an electric field between th...\n",
+      "------------------------------------------------------------\n",
+      "Asking LM Studio...\n",
+      "LLM output invalid JSON: <|channel|>final <|constrain|>JSON<|message|>{\"suggestion\":0,\"confidence\":\"High\",\"rationale\":\"The claim describes a particle acceleration method unrelated to greenhouse gas reduction.\"}\n",
+      "LLM Failed (Check settings or Label Manually)\n",
+      "Saved.\n",
+      "\n",
+      "[Row 50] (Uncertainty: 0.9976)\n",
+      "CLAIM: 1. A biogenic flocculant composition for CEPT sludge conditioning comprising a) a first flocculant component which comprises at least one acidophilic auto-trophic iron-oxidizing bacterium and at least one species of acid tolerant organotrophic microbes which are grown in medium containing iron (II) ...\n",
+      "------------------------------------------------------------\n",
+      "LLM Says: 0 (Medium) | Rationale: The claim focuses on sludge conditioning using microbial flocculants, which is a wastewater treatment application rather than a direct climate‑change mitigation technology.\n",
+      "Saved.\n",
+      "\n",
+      "[Row 51] (Uncertainty: 0.9975)\n",
+      "CLAIM: 1. A nuclear reactor comprising: an elongated reactor vessel enclosed at a lower end and having an open upper end on which an annular flange is formed and a central axis extending, along an elongated dimension; a reactor vessel head having an annular portion on an underside of the bead that is machi...\n",
+      "------------------------------------------------------------\n",
+      "LLM Says: 0 (High) | Rationale: The claim describes a nuclear reactor component, not a climate‑change mitigation technology.\n",
+      "Saved.\n",
+      "\n",
+      "[Row 52] (Uncertainty: 0.9975)\n",
+      "CLAIM: 1. A steam reforming system comprising: a) a kiln, comprising a susceptor tube; a kiln inlet for receiving a feedstock; a conveyor for transporting said feedstock through said kiln; b) a steam reforming reactor comprising a reformer tube; a reactor inlet in fluid communication with said first kiln o...\n",
+      "------------------------------------------------------------\n",
+      "LLM Says: 0 (Medium) | Rationale: The claim describes a steam reforming system for gas production, which is a general chemical process and does not explicitly address greenhouse gas reduction or climate change mitigation.\n",
+      "Saved.\n",
+      "\n",
+      "[Row 53] (Uncertainty: 0.9975)\n",
+      "CLAIM: 1. A pest trap reporting system, comprising: a plurality of pest traps, wherein each pest trap encloses, retains or kills one or more non-human pests; a pest report database that includes pest activity information for the plurality of pest traps; a plurality of sensors, each of the plurality of sens...\n",
+      "------------------------------------------------------------\n",
+      "LLM Says: 0 (High) | Rationale: The claim describes a pest monitoring system, not related to greenhouse gas mitigation.\n",
+      "Saved.\n",
+      "\n",
+      "[Row 54] (Uncertainty: 0.9974)\n",
+      "CLAIM: 1. A motor vehicle comprising: a body; a wheel rotatably supported on the body; an occupant riding portion supported by the body for tilting relative to the body and mounted with an occupant; occupant attitude detection means for detecting an attitude of the occupant riding portion; body attitude de...\n",
+      "------------------------------------------------------------\n",
+      "LLM Says: 0 (High) | Rationale: The claim describes a vehicle tilt-control system unrelated to greenhouse‑gas reduction or climate mitigation.\n",
+      "Saved.\n",
+      "\n",
+      "[Row 55] (Uncertainty: 0.9973)\n",
+      "CLAIM: 1. A compound having Formula (III) or a therapeutically acceptable salt thereof, wherein...\n",
+      "------------------------------------------------------------\n",
+      "Asking LM Studio...\n",
+      "LLM output invalid JSON: <|channel|>final <|constrain|>JSON<|message|>{\"suggestion\":0,\"confidence\":\"High\",\"rationale\":\"The claim describes a chemical compound for therapeutic use, not related to green or climate change mitigation.\"}\n",
+      "LLM Failed (Check settings or Label Manually)\n",
+      "Saved.\n",
+      "\n",
+      "[Row 56] (Uncertainty: 0.9973)\n",
+      "CLAIM: 1. A method for fabricating a semiconductor device, comprising: forming a conductive layer over first and second regions of a semiconductor substrate; forming a trench extended in the first region of the semiconductor substrate through the conductive layer; forming a first gate electrode in the tren...\n",
+      "------------------------------------------------------------\n",
+      "LLM Says: 0 (High) | Rationale: The claim describes semiconductor fabrication steps, unrelated to Green/Climate Change mitigation.\n",
+      "Saved.\n",
+      "\n",
+      "[Row 57] (Uncertainty: 0.9972)\n",
+      "CLAIM: 1. A method for installation of an offshore wind turbine, characterized in comprising the steps of: prefabrication of a foundation, including: fabricating the foundation which includes a plurality of tanks providing buoyant force and uprighting force to the foundation so as to keep the foundation up...\n",
+      "------------------------------------------------------------\n",
+      "LLM Says: 1 (High) | Rationale: The claim describes a method for installing an offshore wind turbine, which is a renewable energy technology that mitigates climate change.\n",
+      "Saved.\n",
+      "\n",
+      "[Row 58] (Uncertainty: 0.9972)\n",
+      "CLAIM: 1. An airfoil comprising: an airfoil body made of a first material with a leading edge, trailing edge, pressure side and suction side; a sheath with first and second flanks made of a second material; a first shim disposed between a portion of an end of the first flank and the airfoil body and extend...\n",
+      "------------------------------------------------------------\n",
+      "LLM Says: 0 (Medium) | Rationale: The claim describes structural components of an airfoil without reference to energy efficiency, emissions reduction, or other climate‑change mitigation measures.\n",
+      "Saved.\n",
+      "\n",
+      "[Row 59] (Uncertainty: 0.9972)\n",
+      "CLAIM: 1. An electric storage system comprising: a plurality of electric storage blocks connected in series, each of the plurality of electric storage blocks including a plurality of electric storage elements connected in parallel; a plurality of current breakers, each of the plurality of current breakers ...\n",
+      "------------------------------------------------------------\n",
+      "LLM Says: 0 (Medium) | Rationale: The claim describes a battery management system, not a direct green or climate change mitigation technology.\n",
+      "Saved.\n",
+      "\n",
+      "[Row 60] (Uncertainty: 0.9972)\n",
+      "CLAIM: 1. A method for performing operations on a stainer in a stainer network comprising: providing a robotic arm coupled to the stainer, the robotic arm having a reagent dispenser; establishing a network connection between a computer and a stainer in the stainer network; sending requests from the compute...\n",
+      "------------------------------------------------------------\n",
+      "LLM Says: 0 (High) | Rationale: The claim describes laboratory automation for sample processing, not a technology related to greenhouse gas mitigation or climate change.\n",
+      "Saved.\n",
+      "\n",
+      "[Row 61] (Uncertainty: 0.9971)\n",
+      "CLAIM: 1. A Group III nitride semiconductor light-emitting device, comprising: a conductive support; a p-electrode disposed on the support; a semiconductor layer disposed on the p-electrode, the semiconductor layer comprising at least a p-layer, a light-emitting layer, and an n-layer disposed in this order...\n",
+      "------------------------------------------------------------\n",
+      "LLM Says: 0 (High) | Rationale: The claim describes a semiconductor light-emitting device, which is unrelated to green or climate change mitigation.\n",
+      "Saved.\n",
+      "\n",
+      "[Row 62] (Uncertainty: 0.9971)\n",
+      "CLAIM: 1. A plate heat exchanger in a sealed design, with: a stacked arrangement comprising: a front-side and a rear-side end plate, wherein at least one end plate is constituted as a connection plate having at least one connection, heat exchanger plates which are arranged and stacked between the front-sid...\n",
+      "------------------------------------------------------------\n",
+      "LLM Says: 0 (High) | Rationale: The claim describes a mechanical design for a plate heat exchanger and does not address energy efficiency or climate‑change mitigation technologies.\n",
+      "Saved.\n",
+      "\n",
+      "[Row 63] (Uncertainty: 0.9971)\n",
+      "CLAIM: 1. A vehicle braking/driving force control system comprising: a braking/driving force generating mechanism that causes each wheel of a vehicle to generate driving force or braking force independently of one another; a suspension mechanism that couples each of the wheels that are not supported by spr...\n",
+      "------------------------------------------------------------\n",
+      "LLM Says: 0 (High) | Rationale: The claim describes a vehicle braking and driving force control system, which is unrelated to greenhouse gas mitigation or climate change technologies.\n",
+      "Saved.\n",
+      "\n",
+      "[Row 64] (Uncertainty: 0.9970)\n",
+      "CLAIM: 1. A detector apparatus configured to receive light and generate electrical signals, the detector apparatus comprising: a light sensor having a light incidence side, the light sensor including at least one photocathode; a cooling component, the cooling component being in direct contact, on the light...\n",
+      "------------------------------------------------------------\n",
+      "LLM Says: 0 (High) | Rationale: The claim describes a light detection device, not related to climate change mitigation.\n",
+      "Saved.\n",
+      "\n",
+      "[Row 65] (Uncertainty: 0.9970)\n",
+      "CLAIM: 1. A flexible display device comprising: a display panel configured to generate an image; and a window member on the display panel, the window member comprising: wherein a width of each of the second parts is smaller than a width of the first part at a bending area....\n",
+      "------------------------------------------------------------\n",
+      "Asking LM Studio...\n",
+      "LLM output invalid JSON: <|channel|>final <|constrain|>JSON<|message|>{\"suggestion\":0,\"confidence\":\"High\",\"rationale\":\"The claim describes a flexible display device, which is unrelated to Green/Climate Change mitigation.\"}\n",
+      "LLM Failed (Check settings or Label Manually)\n",
+      "Saved.\n",
+      "\n",
+      "[Row 66] (Uncertainty: 0.9969)\n",
+      "CLAIM: 1. A conductive film comprising: a substrate; a transparent electrode layer provided on the substrate; and a conductive pattern layer provided on the transparent electrode layer, wherein the conductive pattern layer includes a metal nitride pattern layer including CuNx, x is a mass ratio of N with r...\n",
+      "------------------------------------------------------------\n",
+      "LLM Says: 0 (High) | Rationale: The claim describes a conductive film for electronic applications, with no indication of greenhouse gas reduction or climate mitigation.\n",
+      "Saved.\n",
+      "\n",
+      "[Row 67] (Uncertainty: 0.9969)\n",
+      "CLAIM: 1. A hydrolysable linker selected from a compound of formula V, VI, VII, and VIII: wherein: R′ and R″ are each independently a C each a is independently an integer from 0 to 6; each b is independently an integer from 1 to 6; each X is independently: each X each Y is independently: each Y each m, n, ...\n",
+      "------------------------------------------------------------\n",
+      "LLM Says: 0 (Low) | Rationale: The claim describes a chemical linker, with no indication of climate‑change mitigation.\n",
+      "Saved.\n",
+      "\n",
+      "[Row 68] (Uncertainty: 0.9968)\n",
+      "CLAIM: 1. A substrate bearing a stack of layers as the back contact in a molybdenum photovoltaic device, said back contact comprising in order from the substrate: a barrier layer comprising at least one of: Si a primer layer; a layer of ZnO; and a layer of molybdenum, wherein the molybdenum is deposited di...\n",
+      "------------------------------------------------------------\n",
+      "LLM Says: 1 (High) | Rationale: The claim describes a photovoltaic device structure, which is directly relevant to solar energy technology for climate change mitigation.\n",
+      "Saved.\n",
+      "\n",
+      "[Row 69] (Uncertainty: 0.9968)\n",
+      "CLAIM: 1. A system comprising: a foil including a leading inlet for fluid to enter, a forward chamber within the foil downstream of the leading inlet, a rearward chamber within the foil downstream of the forward chamber, and a constriction formed by the foil between the forward and rearward chambers; at le...\n",
+      "------------------------------------------------------------\n",
+      "LLM Says: 0 (High) | Rationale: The claim describes a fluid flow device without any explicit reference to environmental or climate‑change mitigation.\n",
+      "Saved.\n",
+      "\n",
+      "[Row 70] (Uncertainty: 0.9967)\n",
+      "CLAIM: 1. A method for passivating a surface of crystalline iron disulfide, comprising: sputtering iron disulfide to form a layer of crystalline iron disulfide on a substrate, wherein the layer has a surface comprising crystal surfaces; and depositing a capping layer of epitaxial zinc sulfide onto the surf...\n",
+      "------------------------------------------------------------\n",
+      "LLM Says: 0 (High) | Rationale: The claim concerns surface passivation of iron disulfide, not a technology for greenhouse gas mitigation or climate change.\n",
+      "Saved.\n",
+      "\n",
+      "[Row 71] (Uncertainty: 0.9967)\n",
+      "CLAIM: 1. A brake control apparatus, comprising: a frictional braking unit configured to generate frictional braking force by supplying operating fluid to a wheel cylinder provided on each wheel of a vehicle to press a frictional member against the wheel; a regenerative braking unit configured to generate ...\n",
+      "------------------------------------------------------------\n",
+      "LLM Says: 0 (High) | Rationale: The claim describes a brake control apparatus for vehicles, which is unrelated to Green/Climate Change mitigation.\n",
+      "Saved.\n",
+      "\n",
+      "[Row 72] (Uncertainty: 0.9966)\n",
+      "CLAIM: 1. A circuit for recording a magnitude of an electrostatic discharge (ESD) event during semiconductor assembly, the circuit comprising: a voltage divider connected between a first potential and a second potential, the voltage divider configured to provide a first node having a discrete voltage level...\n",
+      "------------------------------------------------------------\n",
+      "LLM Says: 0 (High) | Rationale: The claim describes an ESD recording circuit for semiconductor manufacturing, which does not pertain to greenhouse gas mitigation or climate change technologies.\n",
+      "Saved.\n",
+      "\n",
+      "[Row 73] (Uncertainty: 0.9966)\n",
+      "CLAIM: 1. A rectifier, comprising: a first rectification unit having an anode and a cathode, the anode being connected to a negative radio frequency (RF) port, and the cathode being connected to a positive direct current (DC) port; a second rectification unit having an anode and a cathode, the anode being ...\n",
+      "------------------------------------------------------------\n",
+      "LLM Says: 0 (High) | Rationale: The claim describes a rectifier circuit, unrelated to climate change mitigation.\n",
+      "Saved.\n",
+      "\n",
+      "[Row 74] (Uncertainty: 0.9966)\n",
+      "CLAIM: 1. An apparatus comprising: a first electronic device to communicate with a second electronic device, the first device comprising:...\n",
+      "------------------------------------------------------------\n",
+      "Asking LM Studio...\n",
+      "LLM output invalid JSON: <|channel|>final <|constrain|>JSON<|message|>{\"suggestion\":0,\"confidence\":\"High\",\"rationale\":\"The claim describes a generic communication apparatus between electronic devices, with no reference to environmental or climate-related functions.\"}\n",
+      "LLM Failed (Check settings or Label Manually)\n",
+      "Saved.\n",
+      "\n",
+      "[Row 75] (Uncertainty: 0.9966)\n",
+      "CLAIM: 1. A process-based method of detecting a CO 2 gas leak in a deep geologic gas storage reservoir, the method comprising: constructing a gas sampling station in a vadose zone proximal to the deep geologic gas storage reservoir; measuring a CO measuring an O measuring a CH measuring a N determining a H...\n",
+      "------------------------------------------------------------\n",
+      "LLM Says: 1 (High) | Rationale: The method detects CO₂ leaks from a geological storage site, directly supporting carbon sequestration efforts.\n",
+      "Saved.\n",
+      "\n",
+      "[Row 76] (Uncertainty: 0.9965)\n",
+      "CLAIM: 1. A self-supporting reflector for a parabolic trough: (a) having a reflectance of at least 90%, based on the solar spectrum; (b) comprising at least one layer of a transparent plastic material facing a light source and having a layer thickness within a range of from 0.1 mm to 8 mm; and (c) at least...\n",
+      "------------------------------------------------------------\n",
+      "LLM Says: 1 (High) | Rationale: The claim describes a component for a parabolic trough solar collector, which is a renewable energy technology used to mitigate climate change.\n",
+      "Saved.\n",
+      "\n",
+      "[Row 77] (Uncertainty: 0.9965)\n",
+      "CLAIM: 1. A valve train system for an internal combustion engine having a combustion chamber with a piston which reciprocates therewithin between a top-dead-center position and a bottom-dead-center position, said valve train system comprising: an intake valve which moves between an intake closed position a...\n",
+      "------------------------------------------------------------\n",
+      "LLM Says: 0 (High) | Rationale: The claim describes a conventional valve train for an internal combustion engine, which is unrelated to Green/Climate Change mitigation.\n",
+      "Saved.\n",
+      "\n",
+      "[Row 78] (Uncertainty: 0.9965)\n",
+      "CLAIM: 1. An organic light-emitting diode (OLED) display, comprising: a first plastic layer; a first barrier layer formed over the first plastic layer; a first intermediate layer formed over the first barrier layer, wherein the first intermediate layer comprises amorphous silicon; a second plastic layer fo...\n",
+      "------------------------------------------------------------\n",
+      "LLM Says: 0 (High) | Rationale: The claim describes an OLED display structure, which does not relate to Green/Climate Change mitigation.\n",
+      "Saved.\n",
+      "\n",
+      "[Row 79] (Uncertainty: 0.9965)\n",
+      "CLAIM: 1. A wave activated power generating device, comprising: a support frame; a buoy vertically positioned to rise and fall relative to motion of waves impacting the buoy and the support frame, the buoy being formed with a hollow interior space; a rack and pinion structure operatively connected between ...\n",
+      "------------------------------------------------------------\n",
+      "LLM Says: 1 (High) | Rationale: The claim describes a wave‑powered generator that converts ocean wave motion into electricity, which is a renewable energy technology for climate change mitigation.\n",
+      "Saved.\n",
+      "\n",
+      "[Row 80] (Uncertainty: 0.9964)\n",
+      "CLAIM: 1. A method for reducing an amount of unwanted living organisms within an algae cultivation fluid, the algae cultivation fluid including wanted living algae of genus Nannochloropsis , the method comprising: subjecting the algae cultivation fluid, the algae cultivation fluid including the wanted livi...\n",
+      "------------------------------------------------------------\n",
+      "Asking LM Studio...\n",
+      "LLM output invalid JSON: <|channel|>final <|constrain|>JSON<|message|>{\"suggestion\":1,\"confidence\":\"Medium\",\"rationale\":\"The\n",
+      "LLM Failed (Check settings or Label Manually)\n",
+      "Saved.\n",
+      "\n",
+      "[Row 81] (Uncertainty: 0.9964)\n",
+      "CLAIM: 1. A compound having Formula (I): wherein, R R b) —(CH R X is —O, NH or S; Y is a cleavable or non-cleavable linker group; and Z is an antigen derived from an infectious agent or a tumor antigen or a pharmaceutically acceptable salt thereof....\n",
+      "------------------------------------------------------------\n",
+      "Asking LM Studio...\n",
+      "LLM output invalid JSON: <|channel|>final <|constrain|>json<|message|>{\"suggestion\":0,\"confidence\":\"High\",\"rationale\":\"The claim concerns a pharmaceutical compound for medical use, not related to climate change mitigation.\"}\n",
+      "LLM Failed (Check settings or Label Manually)\n",
+      "Saved.\n",
+      "\n",
+      "[Row 82] (Uncertainty: 0.9964)\n",
+      "CLAIM: 1. A memory controller comprising: driver circuitry to output a first timing signal to a memory device, the first timing signal to time transmission of a data signal from the memory device to the memory controller; control circuitry to enable oscillation of the first timing signal at a first frequen...\n",
+      "------------------------------------------------------------\n",
+      "LLM Says: 0 (High) | Rationale: The claim describes a memory controller’s timing and clock circuitry, unrelated to climate change mitigation.\n",
+      "Saved.\n",
+      "\n",
+      "[Row 83] (Uncertainty: 0.9964)\n",
+      "CLAIM: 1. A rubber composition, comprising, based on 100% by mass of a rubber component: 5 to 55% by mass of a copolymer (A) of an aromatic vinyl compound and a conjugated diene compound, the copolymer (A) having an aromatic vinyl compound content of 5-14% by mass and a vinyl bond content in the conjugated...\n",
+      "------------------------------------------------------------\n",
+      "LLM Says: 0 (High) | Rationale: The claim describes a rubber composition and silica filler, not a technology for greenhouse gas mitigation.\n",
+      "Saved.\n",
+      "\n",
+      "[Row 84] (Uncertainty: 0.9963)\n",
+      "CLAIM: 1. A method of preparing a metal suboxide, comprising: preparing a mixture including a metal suboxide precursor, an aromatic compound substituted with a hydroxy group, and a linking precursor including one selected from a C1 to C30 aldehyde, a C3 to C30 ketone, and a combination thereof; reacting th...\n",
+      "------------------------------------------------------------\n",
+      "LLM Says: 0 (Low) | Rationale: The claim describes a chemical synthesis method, not directly related to climate change mitigation.\n",
+      "Saved.\n",
+      "\n",
+      "[Row 85] (Uncertainty: 0.9963)\n",
+      "CLAIM: 1. A vehicle hydraulic control device including: an oil pump that is driven by a driving force source for wheels; and an oil passage that guides oil discharged from the oil pump to a rotating electrical machine that forms at least a part of the driving force source and a gear mechanism to which driv...\n",
+      "------------------------------------------------------------\n",
+      "LLM Says: 0 (Medium) | Rationale: The claim describes a hydraulic control system for a vehicle, which does not directly address greenhouse gas reduction or climate change mitigation.\n",
+      "Saved.\n",
+      "\n",
+      "[Row 86] (Uncertainty: 0.9963)\n",
+      "CLAIM: 1. A method for the production of a human or animal nutrition product comprising producing an adsorbate suitable for human or animal nutrition comprising applying a component to a carrier using at least one stabilizer such that the component is adsorbed to the carrier, wherein the carrier has a mean...\n",
+      "------------------------------------------------------------\n",
+      "LLM Says: 0 (High) | Rationale: The claim concerns a nutritional adsorbate production method, not a technology for greenhouse gas reduction or climate mitigation.\n",
+      "Saved.\n",
+      "\n",
+      "[Row 87] (Uncertainty: 0.9963)\n",
+      "CLAIM: 1. An automated driving system, comprising: one or more sensors disposed on an autonomous vehicle; and a computing device in communication with the one or more sensors, comprising:...\n",
+      "------------------------------------------------------------\n",
+      "LLM Says: 0 (Low) | Rationale: The claim describes a general automated driving system without explicit reference to environmental or climate mitigation.\n",
+      "Saved.\n",
+      "\n",
+      "[Row 88] (Uncertainty: 0.9963)\n",
+      "CLAIM: 1. A memory device, comprising: an array of memory cells, the memory cells in the array being programmable to at least two different charge levels; and a control logic unit coupled to the array of memory cells and configured to program the memory cells in each of a plurality of groups with a respect...\n",
+      "------------------------------------------------------------\n",
+      "LLM Says: 0 (High) | Rationale: The claim describes a memory device and programming logic, unrelated to climate change mitigation.\n",
+      "Saved.\n",
+      "\n",
+      "[Row 89] (Uncertainty: 0.9963)\n",
+      "CLAIM: 1. An adjustable solar panel mounting assembly comprising: a. a first clamp further comprising an upper and lower portion wherein the lower portion further comprises a cavity; b. a first mounting plate extending outward from the lower portion of the first clamp to an end; c. a first flange extending...\n",
+      "------------------------------------------------------------\n",
+      "LLM Says: 1 (High) | Rationale: The claim describes a solar panel mounting assembly, which supports renewable energy generation.\n",
+      "Saved.\n",
+      "\n",
+      "[Row 90] (Uncertainty: 0.9963)\n",
+      "CLAIM: 1. A method of planting or seeding multiple types of seed in a single planting pass during row-crop planting or seeding of an agricultural field with an agricultural implement, the method comprising: storing seeds of multiple types including at least a first type and a second type in multiple compar...\n",
+      "------------------------------------------------------------\n",
+      "LLM Says: 0 (Medium) | Rationale: The claim describes a multi‑seed planting method, which is an agricultural technique but does not directly address greenhouse gas reduction or climate change mitigation.\n",
+      "Saved.\n",
+      "\n",
+      "[Row 91] (Uncertainty: 0.9963)\n",
+      "CLAIM: 1. A method to update a cache in a multi-core processor, the method comprising: receiving a notification of a cache miss associated with a process or a thread running on a single core of the multi-core processor, the single core including: determining that an address associated with the cache miss c...\n",
+      "------------------------------------------------------------\n",
+      "LLM Says: 0 (High) | Rationale: The claim concerns processor cache management, unrelated to climate change mitigation.\n",
+      "Saved.\n",
+      "\n",
+      "[Row 92] (Uncertainty: 0.9962)\n",
+      "CLAIM: 1. A driving method of a liquid crystal display device comprising a liquid crystal element, the driving method comprising the steps of: applying a first voltage to the liquid crystal element in a first subframe period of a first frame period; making transmittance of the liquid crystal element at the...\n",
+      "------------------------------------------------------------\n",
+      "LLM Says: 0 (High) | Rationale: The claim describes a technical method for controlling liquid crystal display operation, with no reference to energy efficiency or climate‑change mitigation.\n",
+      "Saved.\n",
+      "\n",
+      "[Row 93] (Uncertainty: 0.9962)\n",
+      "CLAIM: 1. A compound of Formula IA: wherein X R R wherein each R each R where Z wherein the alkyl, alkenyl, alkynyl, cycloalkyl, aryl, heterocyclic, or heteroaryl groups of Z wherein Y indicates one or more optional double bonds; and n is 0, 1, 2, or 3; R each R wherein the alkyl, alkenyl, alkynyl, cycloal...\n",
+      "------------------------------------------------------------\n",
+      "LLM Says: 0 (High) | Rationale: The claim describes a generic chemical structure for a potential pharmaceutical compound with no reference to environmental or climate‑change mitigation.\n",
+      "Saved.\n",
+      "\n",
+      "[Row 94] (Uncertainty: 0.9962)\n",
+      "CLAIM: 1. A valve for a fuel cell comprising: a housing; a first pressure chamber and a second pressure chamber provided in the housing; two supply/discharge tubes connected to the housing, and supplying and discharging fluid to and from the first pressure chamber and the second pressure chamber, respectiv...\n",
+      "------------------------------------------------------------\n",
+      "LLM Says: 1 (Medium) | Rationale: The claim describes a valve for a fuel cell, which is a technology used in clean energy generation and thus relates to green/climate change mitigation.\n",
+      "Saved.\n",
+      "\n",
+      "[Row 95] (Uncertainty: 0.9962)\n",
+      "CLAIM: 1. A wing comprising: an airfoil section including a leading edge, a trailing edge, an upper surface and a lower surface, wherein a region within the airfoil section immediately adjacent the leading edge is ventilated via one or more vent openings which open in the upper surface to establish a sub-s...\n",
+      "------------------------------------------------------------\n",
+      "LLM Says: 0 (High) | Rationale: The claim describes a wing ventilation system for aerodynamic performance, not climate change mitigation.\n",
+      "Saved.\n",
+      "\n",
+      "[Row 96] (Uncertainty: 0.9962)\n",
+      "CLAIM: 1. An input device comprising: a touch panel that includes M (where M is a natural number of 5 or more) driving electrodes, and a plurality of detection electrodes forming capacitances between the respective driving electrodes, in which the M driving electrodes and the plurality of detection electro...\n",
+      "------------------------------------------------------------\n",
+      "LLM Says: 0 (High) | Rationale: The claim describes a touch panel input device, unrelated to green or climate‑change mitigation.\n",
+      "Saved.\n",
+      "\n",
+      "[Row 97] (Uncertainty: 0.9961)\n",
+      "CLAIM: 1. An information display system in a transportation apparatus, the information display system comprises: a liquid crystal display (LCD) screen that occupies at least a portion of a dashboard of the transportation apparatus, wherein the LCD screen is capable of graphically displaying multiple inform...\n",
+      "------------------------------------------------------------\n",
+      "Asking LM Studio...\n",
+      "LLM output invalid JSON: <|channel|>final <|constrain|>JSON<|message|>{\"suggestion\":0,\"confidence\":\"High\",\"rationale\":\"The claim describes an information display system for a vehicle dashboard, which does not directly address greenhouse gas reduction or climate change mitigation.\"}\n",
+      "LLM Failed (Check settings or Label Manually)\n",
+      "Saved.\n",
+      "\n",
+      "[Row 98] (Uncertainty: 0.9960)\n",
+      "CLAIM: 1. A photovoltaic (PV) device, comprising: at least one PV interband cascade (IC PV) stage having a conduction band and a valence band, comprising: wherein the absorption region is positioned between the intraband transport region and the interband tunneling region, wherein the interband tunneling r...\n",
+      "------------------------------------------------------------\n",
+      "LLM Says: 1 (High) | Rationale: The claim describes a photovoltaic device for generating electricity from light.\n",
+      "Saved.\n",
+      "\n",
+      "[Row 99] (Uncertainty: 0.9960)\n",
+      "CLAIM: 1. A rack system comprising: a plurality of trays configured to hold a respective plurality of battery-powered unmanned aerial vehicles; and a frame configured to support the plurality of trays in a vertical arrangement, wherein each tray of the plurality of trays comprises:...\n",
+      "------------------------------------------------------------\n",
+      "Asking LM Studio...\n",
+      "LLM output invalid JSON: <|channel|>final <|constrain|>JSON<|message|>{\"suggestion\":0,\"confidence\":\"Medium\",\"rationale\":\"The claim describes a storage rack for battery-powered UAVs, which does not directly address green or climate change mitigation technologies.\"}\n",
+      "LLM Failed (Check settings or Label Manually)\n",
+      "Saved.\n",
+      "\n",
+      "[Row 100] (Uncertainty: 0.9959)\n",
+      "CLAIM: 1. A DC electrical machine comprising: an armature having a non-integer number of winding slots per pole-pair of a magnetic field of a field means, each winding slot having a phase angle, wherein the phase angle is electrical and is a position of the winding slot in relation to a fundamental wavefor...\n",
+      "------------------------------------------------------------\n",
+      "LLM Says: 0 (High) | Rationale: The claim describes a technical improvement to a DC electrical machine, unrelated to climate change mitigation.\n",
+      "Saved.\n",
+      "\n",
+      "Done\n"
+     ]
+    }
+   ],
+   "source": [
+    "import requests\n",
+    "\n",
+    "# LLM model is locally hosted via network via LM Studio.\n",
+    "LM_STUDIO_URL = \"http://localhost:1234/v1/chat/completions\"\n",
+    "\n",
+    "# GPT-OSS-20B is run locally\n",
+    "MODEL_NAME = \"local-model\" \n",
+    "\n",
+    "filename = \"hitl_green_100.csv\"\n",
+    "\n",
+    "def get_llm_response_lmstudio(claim_text):\n",
+    "    \"\"\"Function to call LM Studio with Error Printing\"\"\"\n",
+    "    \n",
+    "    system_prompt = \"\"\"\n",
+    "    You are a patent classification AI. You must respond in valid JSON format only.\n",
+    "    Schema:\n",
+    "    {\n",
+    "        \"suggestion\": 0 or 1,\n",
+    "        \"confidence\": \"Low\", \"Medium\", or \"High\",\n",
+    "        \"rationale\": \"short sentence\"\n",
+    "    }\n",
+    "    \"\"\"\n",
+    "    \n",
+    "    user_prompt = f\"\"\"\n",
+    "    Analyze this patent claim. Is it related to Green/Climate Change mitigation (Y02)?\n",
+    "    Claim: \"{claim_text[:2000]}\"\n",
+    "    \"\"\"\n",
+    "\n",
+    "    payload = {\n",
+    "        \"model\": MODEL_NAME,\n",
+    "        \"messages\": [\n",
+    "            {\"role\": \"system\", \"content\": system_prompt},\n",
+    "            {\"role\": \"user\", \"content\": user_prompt}\n",
+    "        ],\n",
+    "        \"temperature\": 0.1, \n",
+    "        \"max_tokens\": 150\n",
+    "    }\n",
+    "\n",
+    "    try:\n",
+    "        response = requests.post(LM_STUDIO_URL, json=payload, headers={\"Content-Type\": \"application/json\"})\n",
+    "        \n",
+    "        #DEBUGGING BLOCK\n",
+    "        if response.status_code == 200:\n",
+    "            result = response.json()\n",
+    "            \n",
+    "            # Check if the server sent an error instead of an answer\n",
+    "            if 'choices' not in result:\n",
+    "                print(f\"\\n LM STUDIO ERROR: {result}\")\n",
+    "                return None\n",
+    "                \n",
+    "            content = result['choices'][0]['message']['content']\n",
+    "            \n",
+    "            # Clean up code blocks if present\n",
+    "            if \"```\" in content:\n",
+    "                content = content.replace(\"```json\", \"\").replace(\"```\", \"\").strip()\n",
+    "            \n",
+    "            try:\n",
+    "                return json.loads(content)\n",
+    "            except json.JSONDecodeError:\n",
+    "                print(f\"\\nLLM output invalid JSON: {content}\")\n",
+    "                return None\n",
+    "        else:\n",
+    "            print(f\"Server Error {response.status_code}: {response.text}\")\n",
+    "            return None\n",
+    "            \n",
+    "    except Exception as e:\n",
+    "        print(f\"Connection Error: {e}\")\n",
+    "        return None\n",
+    "\n",
+    "def labeling_loop():\n",
+    "    if not os.path.exists(filename):\n",
+    "        print(f\"Error: {filename} not found.\")\n",
+    "        return\n",
+    "    \n",
+    "    df = pd.read_csv(filename)\n",
+    "    \n",
+    "    # Create columns if missing\n",
+    "    for col in ['llm_green_suggested', 'llm_confidence', 'llm_rationale', 'is_green_human', 'notes']:\n",
+    "        if col not in df.columns: df[col] = \"\"\n",
+    "\n",
+    "    # Find empty rows\n",
+    "    remaining_indices = df[df['is_green_human'].isna() | (df['is_green_human'] == \"\")].index.tolist()\n",
+    "    \n",
+    "    print(f\"--- Part C: HITL Labeling ---\")\n",
+    "    print(f\"Remaining rows: {len(remaining_indices)}\")\n",
+    "    print(\"-----------------------------\\n\")\n",
+    "\n",
+    "    for idx in remaining_indices:\n",
+    "        row = df.loc[idx]\n",
+    "        claim_text = str(row['text'])\n",
+    "        \n",
+    "        print(f\"[Row {idx+1}] (Uncertainty: {row['u']:.4f})\")\n",
+    "        print(f\"CLAIM: {claim_text[:300]}...\") \n",
+    "        print(\"-\" * 60)\n",
+    "\n",
+    "        # 1. Ask LLM\n",
+    "        print(\"Asking LM Studio...\", end=\"\\r\")\n",
+    "        llm_result = get_llm_response_lmstudio(claim_text)\n",
+    "        \n",
+    "        suggestion = 0\n",
+    "        conf = \"Low\"\n",
+    "        rat = \"\"\n",
+    "        \n",
+    "        if llm_result:\n",
+    "            suggestion = llm_result.get('suggestion', 0)\n",
+    "            conf = llm_result.get('confidence', \"Low\")\n",
+    "            rat = llm_result.get('rationale', \"\")\n",
+    "            print(f\"LLM Says: {suggestion} ({conf}) | Rationale: {rat}\")\n",
+    "        else:\n",
+    "            print(\"LLM Failed (Check settings or Label Manually)\")\n",
+    "\n",
+    "        # 2. Human Review\n",
+    "        while True:\n",
+    "            user_input = input(f\"Your Label (0/1) [Enter for {suggestion}]: \")\n",
+    "            if user_input.strip() == \"\":\n",
+    "                final_label = suggestion\n",
+    "                break\n",
+    "            if user_input.strip() in ['0', '1']:\n",
+    "                final_label = int(user_input)\n",
+    "                break\n",
+    "            print(\"Please enter 0 or 1.\")\n",
+    "\n",
+    "        # 3. Save\n",
+    "        df.at[idx, 'llm_green_suggested'] = suggestion\n",
+    "        df.at[idx, 'llm_confidence'] = conf\n",
+    "        df.at[idx, 'llm_rationale'] = rat\n",
+    "        df.at[idx, 'is_green_human'] = final_label\n",
+    "        \n",
+    "        df.to_csv(filename, index=False)\n",
+    "        print(\"Saved.\\n\")\n",
+    "\n",
+    "    print(\"Done\")\n",
+    "\n",
+    "labeling_loop()"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "id": "42e180f2",
+   "metadata": {},
+   "outputs": [
+    {
+     "name": "stdout",
+     "output_type": "stream",
+     "text": [
+      "============================================================\n",
+      "HITL ANALYSIS REPORT\n",
+      "============================================================\n",
+      "Total Claims Labeled: 100\n",
+      "Human Overrides:      3\n",
+      "Agreement Rate:       97.0%\n",
+      "------------------------------------------------------------\n",
+      "\n",
+      " 3 EXAMPLES OF HUMAN OVERRIDES:\n",
+      "\n",
+      "Example #1:\n",
+      "   • Claim Snippet: \"1. An apparatus, comprising: a single, dilute solids phase reactor having a top, a central section, and a bottom section with an exit port, and a top ...\"\n",
+      "   • LLM Suggestion: 0 (Rationale: The claim describes a particle removal apparatus for exhaust gases, which addresses air pollution control rather than greenhouse gas mitigation.)\n",
+      "   • Human Label:    1\n",
+      "   • Your Notes:     Manual override: This technology is classified as Green under CPC Y02.\n",
+      "\n",
+      "Example #2:\n",
+      "   • Claim Snippet: \"1. A biogenic flocculant composition for CEPT sludge conditioning comprising a) a first flocculant component which comprises at least one acidophilic ...\"\n",
+      "   • LLM Suggestion: 0 (Rationale: The claim focuses on sludge conditioning using microbial flocculants, which is a wastewater treatment application rather than a direct climate‑change mitigation technology.)\n",
+      "   • Human Label:    1\n",
+      "   • Your Notes:     Manual override: This technology is classified as Green under CPC Y02.\n",
+      "\n",
+      "Example #3:\n",
+      "   • Claim Snippet: \"1. A nuclear reactor comprising: an elongated reactor vessel enclosed at a lower end and having an open upper end on which an annular flange is formed...\"\n",
+      "   • LLM Suggestion: 0 (Rationale: The claim describes a nuclear reactor component, not a climate‑change mitigation technology.)\n",
+      "   • Human Label:    1\n",
+      "   • Your Notes:     Manual override: This technology is classified as Green under CPC Y02.\n"
+     ]
+    }
+   ],
+   "source": [
+    "# Load the completed file\n",
+    "filename = \"hitl_green_100.csv\"\n",
+    "try:\n",
+    "    df = pd.read_csv(filename)\n",
+    "except FileNotFoundError:\n",
+    "    print(f\"Error: Could not find {filename}. Make sure you saved your work!\")\n",
+    "    exit()\n",
+    "\n",
+    "# Find Disagreements between LLM and Human Labels (0/1)\n",
+    "df['llm_green_suggested'] = pd.to_numeric(df['llm_green_suggested'], errors='coerce').fillna(-1).astype(int)\n",
+    "df['is_green_human'] = pd.to_numeric(df['is_green_human'], errors='coerce').fillna(-1).astype(int)\n",
+    "\n",
+    "overrides = df[df['llm_green_suggested'] != df['is_green_human']]\n",
+    "total_count = len(df)\n",
+    "override_count = len(overrides)\n",
+    "\n",
+    "#Print the Report\n",
+    "print(\"=\"*60)\n",
+    "print(\"HITL ANALYSIS REPORT\")\n",
+    "print(\"=\"*60)\n",
+    "print(f\"Total Claims Labeled: {total_count}\")\n",
+    "print(f\"Human Overrides:      {override_count}\")\n",
+    "print(f\"Agreement Rate:       {((total_count - override_count)/total_count)*100:.1f}%\")\n",
+    "print(\"-\" * 60)\n",
+    "\n",
+    "if override_count > 0:\n",
+    "    print(\"\\n 3 EXAMPLES OF HUMAN OVERRIDES:\")\n",
+    "    # Selecting 3 examples to show\n",
+    "    examples = overrides.head(3)\n",
+    "    \n",
+    "    for i, (idx, row) in enumerate(examples.iterrows(), 1):\n",
+    "        print(f\"\\nExample #{i}:\")\n",
+    "        print(f\"   • Claim Snippet: \\\"{str(row['text'])[:150]}...\\\"\")\n",
+    "        print(f\"   • LLM Suggestion: {row['llm_green_suggested']} (Rationale: {row['llm_rationale']})\")\n",
+    "        print(f\"   • Human Label:    {row['is_green_human']}\")\n",
+    "        if row['notes']:\n",
+    "            print(f\"   • Your Notes:     {row['notes']}\")"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "id": "a4fefb37",
+   "metadata": {},
+   "source": [
+    "# Part D"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "id": "0ad0f2e4",
+   "metadata": {},
+   "outputs": [
+    {
+     "name": "stdout",
+     "output_type": "stream",
+     "text": [
+      "Starting Part D: Final Active Learning Evaluation...\n",
+      "   - Generating Base Training and Eval Embeddings...\n"
+     ]
+    },
+    {
+     "data": {
+      "application/vnd.jupyter.widget-view+json": {
+       "model_id": "213f76fffed1405582e4c06644c3a2cf",
+       "version_major": 2,
+       "version_minor": 0
+      },
+      "text/plain": [
+       "Encoding:   0%|          | 0/63 [00:00<?, ?it/s]"
+      ]
+     },
+     "metadata": {},
+     "output_type": "display_data"
+    },
+    {
+     "data": {
+      "application/vnd.jupyter.widget-view+json": {
+       "model_id": "5c3e96ca30f94a0fa27f27b24368d491",
+       "version_major": 2,
+       "version_minor": 0
+      },
+      "text/plain": [
+       "Encoding:   0%|          | 0/157 [00:00<?, ?it/s]"
+      ]
+     },
+     "metadata": {},
+     "output_type": "display_data"
+    },
+    {
+     "name": "stdout",
+     "output_type": "stream",
+     "text": [
+      "   - Loading 100 human-labeled examples...\n"
+     ]
+    },
+    {
+     "data": {
+      "application/vnd.jupyter.widget-view+json": {
+       "model_id": "21f8b1a47ab249c8bea9de13cd2b80e9",
+       "version_major": 2,
+       "version_minor": 0
+      },
+      "text/plain": [
+       "Encoding:   0%|          | 0/4 [00:00<?, ?it/s]"
+      ]
+     },
+     "metadata": {},
+     "output_type": "display_data"
+    },
+    {
+     "name": "stdout",
+     "output_type": "stream",
+     "text": [
+      "\n",
+      "========================================\n",
+      "FINAL PERFORMANCE COMPARISON\n",
+      "========================================\n",
+      "Metric          | Baseline (Part A)    | Active (Part D)     \n",
+      "------------------------------------------------------------\n",
+      "Precision       |               0.7489 |               0.7473 (-0.0015)\n",
+      "Recall          |               0.7488 |               0.7467 (-0.0021)\n",
+      "F1-score        |               0.7488 |               0.7465 (-0.0023)\n",
+      "============================================================\n",
+     ]
+    }
+   ],
+   "source": [
+    "print(\"Starting Part D: Final Active Learning Evaluation...\")\n",
+    "\n",
+    "# Setup Model & Data\n",
+    "model_name = \"AI-Growth-Lab/PatentSBERTa\"\n",
+    "tokenizer = AutoTokenizer.from_pretrained(model_name)\n",
+    "model = AutoModel.from_pretrained(model_name)\n",
+    "device = torch.device(\"cuda\" if torch.cuda.is_available() else \"cpu\")\n",
+    "model.to(device); model.eval()\n",
+    "\n",
+    "def get_embeddings(text_list, batch_size=32):\n",
+    "    all_embeddings = []\n",
+    "    for i in tqdm(range(0, len(text_list), batch_size), desc=\"Encoding\"):\n",
+    "        batch_texts = text_list[i:i+batch_size]\n",
+    "        inputs = tokenizer(batch_texts, padding=True, truncation=True, max_length=128, return_tensors=\"pt\").to(device)\n",
+    "        with torch.no_grad():\n",
+    "            outputs = model(**inputs)\n",
+    "            all_embeddings.append(outputs.last_hidden_state[:, 0, :].cpu().numpy())\n",
+    "    return np.vstack(all_embeddings)\n",
+    "\n",
+    "# Re-create splits from Part A\n",
+    "df = pd.read_parquet(\"patents_50k_green.parquet\")\n",
+    "df_eval = df.sample(n=5000, random_state=42)\n",
+    "df_remaining = df.drop(df_eval.index)\n",
+    "df_train = df_remaining.sample(n=2000, random_state=42)\n",
+    "\n",
+    "print(\"   - Generating Base Training and Eval Embeddings...\")\n",
+    "X_train = get_embeddings(df_train['text'].tolist())\n",
+    "y_train = df_train['is_green_silver'].values\n",
+    "X_eval = get_embeddings(df_eval['text'].tolist())\n",
+    "y_eval = df_eval['is_green_silver'].values\n",
+    "\n",
+    "# Train Baseline\n",
+    "clf_base = LogisticRegression(max_iter=1000, random_state=42)\n",
+    "clf_base.fit(X_train, y_train)\n",
+    "base_report = classification_report(y_eval, clf_base.predict(X_eval), output_dict=True)\n",
+    "\n",
+    "# Load your HITL Gold Labels\n",
+    "df_hitl = pd.read_csv(\"hitl_green_100.csv\")\n",
+    "print(f\"   - Loading {len(df_hitl)} human-labeled examples...\")\n",
+    "X_hitl = get_embeddings(df_hitl['text'].tolist())\n",
+    "y_hitl = df_hitl['is_green_human'].values\n",
+    "\n",
+    "# Active Learning: Combine Original Train + Human Gold Labels\n",
+    "X_combined = np.vstack([X_train, X_hitl])\n",
+    "y_combined = np.concatenate([y_train, y_hitl])\n",
+    "\n",
+    "# Train the Active Learning Model\n",
+    "clf_active = LogisticRegression(max_iter=1000, random_state=42)\n",
+    "clf_active.fit(X_combined, y_combined)\n",
+    "active_report = classification_report(y_eval, clf_active.predict(X_eval), output_dict=True)\n",
+    "\n",
+    "# FINAL COMPARISON REPORT\n",
+    "print(\"\\n\" + \"=\"*40)\n",
+    "print(\"FINAL PERFORMANCE COMPARISON\")\n",
+    "print(\"=\"*40)\n",
+    "print(f\"{'Metric':<15} | {'Baseline (Part A)':<20} | {'Active (Part D)':<20}\")\n",
+    "print(\"-\" * 60)\n",
+    "for m in ['precision', 'recall', 'f1-score']:\n",
+    "    val_a = base_report['macro avg'][m]\n",
+    "    val_d = active_report['macro avg'][m]\n",
+    "    diff = val_d - val_a\n",
+    "    print(f\"{m.capitalize():<15} | {val_a:20.4f} | {val_d:20.4f} ({'+' if diff >=0 else ''}{diff:.4f})\")\n",
+    "print(\"=\"*60)\n",
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "id": "cf420b77",
+   "metadata": {},
+   "outputs": [],
+   "source": []
+  }
+ ],
+ "metadata": {
+  "kernelspec": {
+   "display_name": "Python 3",
+   "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": "3.12.10"
+  }
+ },
+ "nbformat": 4,
+ "nbformat_minor": 5
+}

patents_50k_green.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:4ce03253c78337876a32fa166356cc6a5b87f66488f3aac55be4c02420fa1fb6
+size 21754536