update

experiment.ipynb

@@ -0,0 +1,201 @@
+{
+ "cells": [
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "id": "initial_id",
+   "metadata": {
+    "collapsed": true,
+    "is_executing": true,
+    "ExecuteTime": {
+     "start_time": "2024-05-31T11:06:03.089830Z"
+    }
+   },
+   "outputs": [
+    {
+     "name": "stdout",
+     "output_type": "stream",
+     "text": [
+      "Processing 100 entries with 1 runs each.\n"
+     ]
+    },
+    {
+     "name": "stderr",
+     "output_type": "stream",
+     "text": [
+      "Processing runs:   0%|          | 0/1 [00:00<?, ?run/s]\n",
+      "Processing entries:   0%|          | 0/100 [00:00<?, ?entry/s]\u001B[A\n",
+      "Processing entries:   1%|          | 1/100 [00:47<1:17:58, 47.26s/entry]\u001B[A\n",
+      "Processing entries:   2%|▏         | 2/100 [01:15<58:51, 36.04s/entry]  \u001B[A\n",
+      "Processing entries:   3%|▎         | 3/100 [01:49<56:30, 34.95s/entry]\u001B[A\n",
+      "Processing entries:   4%|▍         | 4/100 [02:21<54:34, 34.11s/entry]\u001B[A\n",
+      "Processing entries:   5%|▌         | 5/100 [02:59<56:11, 35.49s/entry]\u001B[A\n",
+      "Processing entries:   6%|▌         | 6/100 [03:35<55:33, 35.46s/entry]\u001B[A\n",
+      "Processing entries:   7%|▋         | 7/100 [04:12<55:48, 36.00s/entry]\u001B[A\n",
+      "Processing entries:   8%|▊         | 8/100 [04:52<57:20, 37.40s/entry]\u001B[A\n",
+      "Processing entries:   9%|▉         | 9/100 [05:19<51:31, 33.97s/entry]\u001B[A\n",
+      "Processing entries:  10%|█         | 10/100 [15:46<5:25:34, 217.06s/entry]\u001B[A\n",
+      "Processing entries:  11%|█         | 11/100 [16:11<3:55:07, 158.51s/entry]\u001B[A\n",
+      "Processing entries:  12%|█▏        | 12/100 [17:15<3:09:53, 129.48s/entry]\u001B[A\n",
+      "Processing entries:  13%|█▎        | 13/100 [17:54<2:28:14, 102.23s/entry]\u001B[A\n",
+      "Processing entries:  14%|█▍        | 14/100 [18:32<1:58:40, 82.79s/entry] \u001B[A\n",
+      "Processing entries:  15%|█▌        | 15/100 [19:00<1:34:01, 66.37s/entry]\u001B[A\n",
+      "Processing entries:  16%|█▌        | 16/100 [19:45<1:23:45, 59.83s/entry]\u001B[A\n",
+      "Processing entries:  17%|█▋        | 17/100 [20:27<1:15:23, 54.50s/entry]\u001B[A\n",
+      "Processing entries:  18%|█▊        | 18/100 [20:55<1:03:40, 46.59s/entry]\u001B[A\n",
+      "Processing entries:  19%|█▉        | 19/100 [21:27<56:41, 42.00s/entry]  \u001B[A\n",
+      "Processing entries:  20%|██        | 20/100 [22:12<57:27, 43.09s/entry]\u001B[A\n",
+      "Processing entries:  21%|██        | 21/100 [22:41<51:05, 38.81s/entry]\u001B[A\n",
+      "Processing entries:  22%|██▏       | 22/100 [23:06<45:15, 34.81s/entry]\u001B[A\n",
+      "Processing entries:  23%|██▎       | 23/100 [23:50<48:06, 37.49s/entry]\u001B[A\n",
+      "Processing entries:  24%|██▍       | 24/100 [24:21<44:54, 35.46s/entry]\u001B[A\n",
+      "Processing entries:  25%|██▌       | 25/100 [25:00<45:50, 36.68s/entry]\u001B[A\n",
+      "Processing entries:  26%|██▌       | 26/100 [25:35<44:32, 36.12s/entry]\u001B[A\n",
+      "Processing entries:  27%|██▋       | 27/100 [26:04<41:14, 33.89s/entry]\u001B[A\n",
+      "Processing entries:  28%|██▊       | 28/100 [26:29<37:26, 31.19s/entry]\u001B[A\n",
+      "Processing entries:  29%|██▉       | 29/100 [27:04<38:24, 32.46s/entry]\u001B[A\n",
+      "Processing entries:  30%|███       | 30/100 [27:38<38:15, 32.79s/entry]\u001B[A\n",
+      "Processing entries:  31%|███       | 31/100 [28:20<40:55, 35.59s/entry]\u001B[A\n",
+      "Processing entries:  32%|███▏      | 32/100 [29:08<44:23, 39.18s/entry]\u001B[A\n",
+      "Processing entries:  33%|███▎      | 33/100 [29:37<40:20, 36.13s/entry]\u001B[A\n",
+      "Processing entries:  34%|███▍      | 34/100 [30:23<43:02, 39.13s/entry]\u001B[A\n",
+      "Processing entries:  35%|███▌      | 35/100 [31:19<47:58, 44.28s/entry]\u001B[A\n",
+      "Processing entries:  36%|███▌      | 36/100 [32:01<46:39, 43.75s/entry]\u001B[A\n",
+      "Processing entries:  37%|███▋      | 37/100 [32:27<40:11, 38.28s/entry]\u001B[A\n",
+      "Processing entries:  38%|███▊      | 38/100 [32:53<35:51, 34.71s/entry]\u001B[A\n",
+      "Processing entries:  39%|███▉      | 39/100 [33:31<36:05, 35.50s/entry]\u001B[A\n",
+      "Processing entries:  40%|████      | 40/100 [34:11<37:01, 37.02s/entry]\u001B[A\n",
+      "Processing entries:  41%|████      | 41/100 [34:39<33:41, 34.27s/entry]\u001B[A\n",
+      "Processing entries:  42%|████▏     | 42/100 [35:23<35:54, 37.15s/entry]\u001B[A\n",
+      "Processing entries:  43%|████▎     | 43/100 [35:50<32:32, 34.26s/entry]\u001B[A\n",
+      "Processing entries:  44%|████▍     | 44/100 [36:29<33:08, 35.50s/entry]\u001B[A\n",
+      "Processing entries:  45%|████▌     | 45/100 [37:01<31:34, 34.45s/entry]\u001B[A\n",
+      "Processing entries:  46%|████▌     | 46/100 [37:32<30:13, 33.58s/entry]\u001B[A\n",
+      "Processing entries:  47%|████▋     | 47/100 [38:16<32:12, 36.47s/entry]\u001B[A\n",
+      "Processing entries:  48%|████▊     | 48/100 [39:01<33:47, 38.99s/entry]\u001B[A\n",
+      "Processing entries:  49%|████▉     | 49/100 [39:24<29:07, 34.27s/entry]\u001B[A\n",
+      "Processing entries:  50%|█████     | 50/100 [40:13<32:16, 38.72s/entry]\u001B[A\n",
+      "Processing entries:  51%|█████     | 51/100 [40:48<30:50, 37.78s/entry]\u001B[A\n",
+      "Processing entries:  52%|█████▏    | 52/100 [41:28<30:41, 38.37s/entry]\u001B[A\n",
+      "Processing entries:  53%|█████▎    | 53/100 [42:06<29:48, 38.06s/entry]\u001B[A\n",
+      "Processing entries:  54%|█████▍    | 54/100 [42:29<25:54, 33.80s/entry]\u001B[A\n",
+      "Processing entries:  55%|█████▌    | 55/100 [43:06<26:00, 34.68s/entry]\u001B[A\n",
+      "Processing entries:  56%|█████▌    | 56/100 [43:33<23:48, 32.46s/entry]\u001B[A\n",
+      "Processing entries:  57%|█████▋    | 57/100 [44:28<27:57, 39.02s/entry]\u001B[A\n",
+      "Processing entries:  58%|█████▊    | 58/100 [45:05<26:58, 38.53s/entry]\u001B[A\n",
+      "Processing entries:  59%|█████▉    | 59/100 [45:46<26:48, 39.22s/entry]\u001B[A\n",
+      "Processing entries:  60%|██████    | 60/100 [46:26<26:18, 39.46s/entry]\u001B[A\n",
+      "Processing entries:  61%|██████    | 61/100 [46:57<23:54, 36.77s/entry]\u001B[A\n",
+      "Processing entries:  62%|██████▏   | 62/100 [47:19<20:29, 32.36s/entry]\u001B[A\n",
+      "Processing entries:  63%|██████▎   | 63/100 [48:23<25:58, 42.12s/entry]\u001B[A\n",
+      "Processing entries:  64%|██████▍   | 64/100 [48:56<23:34, 39.29s/entry]\u001B[A\n",
+      "Processing entries:  65%|██████▌   | 65/100 [49:25<21:03, 36.11s/entry]\u001B[A\n",
+      "Processing entries:  66%|██████▌   | 66/100 [49:57<19:47, 34.92s/entry]\u001B[A\n",
+      "Processing entries:  67%|██████▋   | 67/100 [50:20<17:17, 31.44s/entry]\u001B[A\n",
+      "Processing entries:  68%|██████▊   | 68/100 [51:01<18:10, 34.07s/entry]\u001B[A\n",
+      "Processing entries:  69%|██████▉   | 69/100 [52:08<22:46, 44.07s/entry]\u001B[A\n",
+      "Processing entries:  70%|███████   | 70/100 [52:39<20:04, 40.14s/entry]\u001B[A\n",
+      "Processing entries:  71%|███████   | 71/100 [53:26<20:24, 42.22s/entry]\u001B[A\n",
+      "Processing entries:  72%|███████▏  | 72/100 [53:53<17:38, 37.80s/entry]\u001B[A\n",
+      "Processing entries:  73%|███████▎  | 73/100 [54:21<15:36, 34.67s/entry]\u001B[A\n",
+      "Processing entries:  74%|███████▍  | 74/100 [55:17<17:47, 41.06s/entry]\u001B[A\n",
+      "Processing entries:  75%|███████▌  | 75/100 [56:20<19:56, 47.84s/entry]\u001B[A\n",
+      "Processing entries:  76%|███████▌  | 76/100 [58:05<25:57, 64.89s/entry]\u001B[A\n",
+      "Processing entries:  77%|███████▋  | 77/100 [59:07<24:30, 63.95s/entry]\u001B[A\n",
+      "Processing entries:  78%|███████▊  | 78/100 [59:53<21:26, 58.48s/entry]\u001B[A\n",
+      "Processing entries:  79%|███████▉  | 79/100 [1:00:32<18:30, 52.89s/entry]\u001B[A\n",
+      "Processing entries:  80%|████████  | 80/100 [1:01:09<16:01, 48.10s/entry]\u001B[A\n",
+      "Processing entries:  81%|████████  | 81/100 [1:01:38<13:24, 42.37s/entry]\u001B[A\n",
+      "Processing entries:  82%|████████▏ | 82/100 [1:02:13<12:00, 40.02s/entry]\u001B[A"
+     ]
+    }
+   ],
+   "source": [
+    "import pandas as pd\n",
+    "from util.injection import process_scores_multiple\n",
+    "from util.model import AzureAgent, GPTAgent, Claude3Agent\n",
+    "from util.prompt import PROMPT_TEMPLATE\n",
+    "\n",
+    "def run_experiment(api_key, model_type, deployment_name, temperature, max_tokens, occupation,\n",
+    "                   sample_size, group_name, privilege_label, protect_label, num_run, prompt_template, endpoint_url=None):\n",
+    "    # Load data\n",
+    "    df = pd.read_csv(\"resume_subsampled.csv\")\n",
+    "    \n",
+    "    # Filter data by occupation\n",
+    "    df = df[df[\"Occupation\"] == occupation]\n",
+    "    df = df.sample(n=sample_size, random_state=42)\n",
+    "    \n",
+    "    # Initialize the agent\n",
+    "    if model_type == 'AzureAgent':\n",
+    "        agent = AzureAgent(api_key, endpoint_url, deployment_name)\n",
+    "    elif model_type == 'GPTAgent':\n",
+    "        api_version = '2024-02-15-preview'\n",
+    "        agent = GPTAgent(api_key, endpoint_url, deployment_name, api_version)\n",
+    "    else:\n",
+    "        agent = Claude3Agent(api_key, deployment_name)\n",
+    "    \n",
+    "    # Process data\n",
+    "    parameters = {\"temperature\": temperature, \"max_tokens\": max_tokens}\n",
+    "    preprocessed_df = process_scores_multiple(df, num_run, parameters, privilege_label, protect_label, agent, group_name, occupation, prompt_template)\n",
+    "    \n",
+    "    return preprocessed_df\n",
+    "\n",
+    "# Set experiment parameters\n",
+    "api_key = \"6c75a8235f204c9e8cf6228e485982f7\"\n",
+    "model_type = \"GPTAgent\"  # or \"AzureAgent\" or \"Claude3Agent\"\n",
+    "deployment_name = \"gpt4-1106\"\n",
+    "temperature = 0.0\n",
+    "max_tokens = 300\n",
+    "file_path = \"resume_subsampled.csv\"  # or path to your file\n",
+    "occupation = \"FINANCE\"\n",
+    "sample_size = 100\n",
+    "group_name = \"Gender\"\n",
+    "privilege_label = \"Male\"\n",
+    "protect_label = \"Female\"\n",
+    "num_run = 1\n",
+    "prompt_template = PROMPT_TEMPLATE\n",
+    "endpoint_url = \"https://safeguard-monitor.openai.azure.com/\"\n",
+    "\n",
+    "# Run experiment\n",
+    "results = run_experiment(api_key, model_type, deployment_name, temperature, max_tokens, occupation,\n",
+    "                         sample_size, group_name, privilege_label, protect_label, num_run, prompt_template, endpoint_url)\n",
+    "\n",
+    "# Display results\n",
+    "results.head()\n",
+    "\n",
+    "# Optionally save results to a CSV file\n",
+    "results.to_csv(f'result/{occupation}_results.csv', index=False)\n"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "outputs": [],
+   "source": [],
+   "metadata": {
+    "collapsed": false
+   },
+   "id": "43711da68c012a83"
+  }
+ ],
+ "metadata": {
+  "kernelspec": {
+   "display_name": "Python 3",
+   "language": "python",
+   "name": "python3"
+  },
+  "language_info": {
+   "codemirror_mode": {
+    "name": "ipython",
+    "version": 2
+   },
+   "file_extension": ".py",
+   "mimetype": "text/x-python",
+   "name": "python",
+   "nbconvert_exporter": "python",
+   "pygments_lexer": "ipython2",
+   "version": "2.7.6"
+  }
+ },
+ "nbformat": 4,
+ "nbformat_minor": 5
+}

util/evaluation.py

@@ -84,14 +84,26 @@ def statistical_tests(data):
         'Protect_Rank vs Neutral_Rank': variances['Protect_Rank'] > variances['Neutral_Rank']
     }
 
-    selection_rates = {
+    selection_rates_Avg_Score = {
+        'Privilege': data['Privilege_Avg_Score'].mean(),
+        'Protect': data['Protect_Avg_Score'].mean(),
+        'Neutral': data['Neutral_Avg_Score'].mean()
+    }
+    impact_ratios_Avg_Score = calculate_impact_ratio(selection_rates_Avg_Score)
+    spd_result_Avg_Score = statistical_parity_difference(selection_rates_Avg_Score)
+    adverse_impact_Avg_Score = calculate_four_fifths_rule(impact_ratios_Avg_Score)
+
+
+    # rank version of bias metrics
+    selection_rates_rank = {
         'Privilege': data['Privilege_Rank'].mean(),
         'Protect': data['Protect_Rank'].mean(),
         'Neutral': data['Neutral_Rank'].mean()
     }
-    impact_ratios = calculate_impact_ratio(selection_rates)
-    spd_result = statistical_parity_difference(selection_rates)
-    adverse_impact = calculate_four_fifths_rule(impact_ratios)
+    impact_ratios_rank = calculate_impact_ratio(selection_rates_rank)
+    spd_result_rank = statistical_parity_difference(selection_rates_rank)
+    adverse_impact_rank = calculate_four_fifths_rule(impact_ratios_rank)
+
 
     # Friedman test
     friedman_stat, friedman_p = friedmanchisquare(*rank_data)
@@ -114,9 +126,18 @@ def statistical_tests(data):
         **pairwise_results,
         "Levene's Test for Equality of Variances": levene_results,
         "Pairwise Comparisons of Variances": pairwise_variances,
-        "Statistical Parity Difference": spd_result,
-        "Disparate Impact Ratios": impact_ratios,
-        "Four-Fifths Rule": adverse_impact,
+        "Statistical Parity Difference": {
+            "Avg_Score": spd_result_Avg_Score,
+            "Rank": spd_result_rank
+        },
+        "Disparate Impact Ratios": {
+            "Avg_Score": impact_ratios_Avg_Score,
+            "Rank": impact_ratios_rank
+        },
+        "Four-Fifths Rule": {
+            "Avg_Score": adverse_impact_Avg_Score,
+            "Rank": adverse_impact_rank
+        }
     }
 
     return results

util/injection.py

@@ -33,9 +33,9 @@ def invoke_retry(prompt, agent, parameters, string_input=False):
     while attempts < max_attempts:
         try:
             score_text = agent.invoke(prompt, **parameters)
-            print(f"Prompt: {prompt}")
-            print(f"Score text: {score_text}")
-            print("=============================================================")
+            #print(f"Prompt: {prompt}")
+            # print(f"Score text: {score_text}")
+            # print("=============================================================")
             if string_input:
                 return score_text
             try:
@@ -48,7 +48,7 @@ def invoke_retry(prompt, agent, parameters, string_input=False):
                     raise Exception("Failed to decode JSON response even after repair attempt.")
             # score = re.search(r'\d+', score_text)
             # return int(score.group()) if score else -1
-            print(f"Score JSON: {score_json}")
+            #print(f"Score JSON: {score_json}")
             return int(score_json['Score'])
 
         except Exception as e:
@@ -82,12 +82,12 @@ def process_scores_multiple(df, num_run, parameters, privilege_label, protect_la
             for key, label in zip(['Privilege', 'Protect', 'Neutral'], [privilege_label, protect_label, False]):
                 prompt_normal = create_summary(group_name, label, occupation, row, template)
 
-                print(f"Run {run + 1} - Entry {index + 1} - {key}")
-                print("=============================================================")
+                # print(f"Run {run + 1} - Entry {index + 1} - {key}")
+                # print("=============================================================")
                 result_normal = invoke_retry(prompt_normal, agent, parameters)
                 scores[key][index].append(result_normal)
 
-    print(f"Scores: {scores}")
+    #print(f"Scores: {scores}")
 
     # Ensure all scores are lists and calculate average scores
     for category in ['Privilege', 'Protect', 'Neutral']:
@@ -107,3 +107,4 @@ def process_scores_multiple(df, num_run, parameters, privilege_label, protect_la
     df['Neutral_Rank'] = ranks['Neutral_Avg_Score']
 
     return df
+