File size: 68,249 Bytes
71c12a0 0788177 71c12a0 0788177 71c12a0 0788177 71c12a0 0788177 71c12a0 0788177 71c12a0 0788177 71c12a0 0788177 71c12a0 0788177 71c12a0 0788177 71c12a0 0788177 71c12a0 0788177 71c12a0 0788177 71c12a0 0788177 71c12a0 0788177 71c12a0 0788177 71c12a0 0788177 71c12a0 0788177 71c12a0 0788177 71c12a0 0788177 71c12a0 0788177 71c12a0 0788177 71c12a0 0788177 71c12a0 0788177 71c12a0 0788177 71c12a0 0788177 71c12a0 0788177 71c12a0 0788177 71c12a0 0788177 71c12a0 0788177 71c12a0 0788177 71c12a0 0788177 71c12a0 0788177 71c12a0 0788177 71c12a0 0788177 71c12a0 0788177 71c12a0 0788177 71c12a0 0788177 71c12a0 0788177 71c12a0 0788177 71c12a0 0788177 71c12a0 0788177 71c12a0 0788177 71c12a0 0788177 71c12a0 0788177 71c12a0 0788177 71c12a0 0788177 71c12a0 0788177 71c12a0 0788177 71c12a0 0788177 71c12a0 0788177 71c12a0 0788177 71c12a0 0788177 71c12a0 0788177 71c12a0 0788177 71c12a0 0788177 71c12a0 0788177 71c12a0 0788177 71c12a0 0788177 71c12a0 0788177 71c12a0 0788177 71c12a0 0788177 71c12a0 0788177 71c12a0 0788177 71c12a0 0788177 71c12a0 0788177 71c12a0 0788177 71c12a0 0788177 71c12a0 0788177 71c12a0 0788177 71c12a0 0788177 71c12a0 9133db0 71c12a0 0788177 71c12a0 9133db0 71c12a0 0788177 71c12a0 |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502 503 504 505 506 507 508 509 510 511 512 513 514 515 516 517 518 519 520 521 522 523 524 525 526 527 528 529 530 531 532 533 534 535 536 537 538 539 540 541 542 543 544 545 546 547 548 549 550 551 552 553 554 555 556 557 558 559 560 561 562 563 564 565 566 567 568 569 570 571 572 573 574 575 576 577 578 579 580 581 582 583 584 585 586 587 588 589 590 591 592 593 594 595 596 597 598 599 600 601 602 603 604 605 606 607 608 609 610 611 612 613 614 615 616 617 618 619 620 621 622 623 624 625 626 627 628 629 630 631 632 633 634 635 636 637 638 639 640 641 642 643 644 645 646 647 648 649 650 651 652 653 654 655 656 657 658 659 660 661 662 663 664 665 666 667 668 669 670 671 672 673 674 675 676 677 678 679 680 681 682 683 684 685 686 687 688 689 690 691 692 693 694 695 696 697 698 699 700 701 702 703 704 705 706 707 708 709 710 711 712 713 714 715 716 717 718 719 720 721 722 723 724 725 726 727 728 729 730 731 732 733 734 735 736 737 738 739 740 741 742 743 744 745 746 747 748 749 750 751 752 753 754 755 756 757 758 759 760 761 762 763 764 765 766 767 768 769 770 771 772 773 774 775 776 777 778 779 780 781 782 783 784 785 786 787 788 789 790 791 792 793 794 795 796 797 798 799 800 801 802 803 804 805 806 807 808 809 810 811 812 813 814 815 816 817 818 819 820 821 822 823 824 825 826 827 828 829 830 831 832 833 834 835 836 837 838 839 840 841 842 843 844 845 846 847 848 849 850 851 852 853 854 855 856 857 858 859 860 861 862 863 864 865 866 867 868 869 870 871 872 873 874 875 876 877 878 879 880 881 882 883 884 885 886 887 888 889 890 891 892 893 894 895 896 897 898 899 900 901 902 903 904 905 906 907 908 909 910 911 912 913 914 915 916 917 918 919 920 921 922 923 924 925 926 927 928 929 930 931 932 933 934 935 936 937 938 939 940 941 942 943 944 945 946 947 948 949 950 951 952 953 954 955 956 957 958 959 960 961 962 963 964 965 966 967 968 969 970 971 972 973 974 975 976 977 978 979 980 981 982 983 984 985 986 987 988 989 990 991 992 993 994 995 996 997 998 999 1000 1001 1002 1003 1004 1005 1006 1007 1008 1009 1010 1011 1012 1013 1014 1015 1016 1017 1018 1019 1020 1021 1022 1023 1024 1025 1026 1027 1028 1029 1030 1031 1032 1033 1034 1035 1036 1037 1038 1039 1040 1041 1042 1043 1044 1045 1046 1047 1048 1049 1050 1051 1052 1053 1054 1055 1056 1057 1058 1059 1060 1061 1062 1063 1064 1065 1066 1067 1068 1069 1070 1071 1072 1073 1074 1075 1076 1077 1078 1079 1080 1081 1082 1083 1084 1085 1086 1087 1088 1089 1090 1091 1092 1093 1094 1095 1096 1097 1098 1099 1100 1101 1102 1103 1104 1105 1106 1107 1108 1109 1110 1111 1112 1113 1114 1115 1116 1117 1118 1119 1120 1121 1122 1123 1124 1125 1126 1127 1128 1129 1130 1131 1132 1133 1134 1135 1136 1137 1138 1139 1140 1141 1142 1143 1144 1145 1146 1147 1148 1149 1150 1151 1152 1153 1154 1155 1156 1157 1158 1159 1160 1161 1162 1163 1164 1165 1166 1167 1168 1169 1170 1171 1172 1173 1174 1175 1176 1177 1178 1179 1180 1181 1182 1183 1184 1185 1186 1187 1188 1189 1190 1191 1192 1193 1194 1195 1196 1197 1198 1199 1200 1201 1202 1203 1204 1205 1206 1207 1208 1209 1210 1211 1212 1213 1214 1215 1216 1217 1218 1219 1220 1221 1222 1223 1224 1225 1226 1227 1228 1229 1230 1231 1232 1233 1234 1235 1236 1237 1238 1239 1240 1241 1242 1243 1244 1245 1246 1247 1248 1249 1250 1251 1252 1253 1254 1255 1256 1257 1258 1259 1260 1261 1262 1263 1264 1265 1266 1267 1268 1269 1270 1271 1272 1273 1274 1275 1276 1277 1278 1279 1280 1281 1282 1283 1284 1285 1286 1287 1288 1289 1290 1291 1292 1293 1294 1295 1296 1297 1298 1299 1300 1301 1302 1303 1304 1305 1306 1307 1308 1309 1310 1311 1312 1313 1314 1315 1316 1317 1318 1319 1320 1321 1322 1323 1324 1325 1326 1327 1328 1329 1330 1331 1332 1333 1334 1335 1336 1337 1338 1339 1340 1341 1342 1343 1344 1345 1346 1347 1348 1349 1350 1351 1352 1353 1354 1355 1356 1357 1358 1359 1360 1361 1362 1363 1364 1365 1366 1367 1368 1369 1370 1371 1372 1373 1374 1375 1376 1377 1378 1379 1380 1381 1382 1383 1384 1385 1386 1387 1388 1389 1390 1391 1392 1393 1394 1395 1396 1397 1398 1399 1400 1401 1402 1403 1404 1405 1406 1407 1408 1409 1410 1411 1412 1413 1414 1415 1416 1417 1418 1419 1420 1421 1422 1423 1424 1425 1426 1427 1428 1429 1430 1431 1432 1433 1434 1435 1436 1437 1438 1439 1440 1441 1442 1443 1444 1445 1446 1447 1448 1449 1450 1451 1452 1453 1454 1455 1456 1457 1458 1459 1460 |
{
"cells": [
{
"cell_type": "code",
"execution_count": 1,
"id": "5f167a6f-5139-46e6-afb2-a1fa4d12f3fd",
"metadata": {},
"outputs": [],
"source": [
"import time\n",
"import logging\n",
"import re\n",
"import random\n",
"import gc\n",
"import numpy as np\n",
"import pandas as pd\n",
"import torch\n",
"import evaluate\n",
"\n",
"from datasets import Dataset, DatasetDict, load_from_disk\n",
"from transformers import (\n",
" AutoModelForSeq2SeqLM,\n",
" AutoTokenizer,\n",
" TrainingArguments,\n",
" Trainer,\n",
" GenerationConfig,\n",
" BitsAndBytesConfig,\n",
")\n",
"from transformers.trainer_callback import EarlyStoppingCallback\n",
"from peft import LoraConfig, get_peft_model, prepare_model_for_kbit_training"
]
},
{
"cell_type": "code",
"execution_count": 2,
"id": "53684b5e-c27e-4eb9-815e-583aa194e096",
"metadata": {},
"outputs": [
{
"name": "stdout",
"output_type": "stream",
"text": [
"cuda\n"
]
}
],
"source": [
"# Enable cudnn benchmark for fixed input sizes (can speed up computation)\n",
"torch.backends.cudnn.benchmark = True\n",
"\n",
"# Set device to RTX 4090\n",
"device = torch.device(\"cuda\" if torch.cuda.is_available() else \"cpu\")\n",
"print(device)"
]
},
{
"cell_type": "code",
"execution_count": 3,
"id": "a47bf3cd-752d-4d1c-9697-70098d6204fa",
"metadata": {},
"outputs": [],
"source": [
"random.seed(42)\n",
"np.random.seed(42)\n",
"torch.manual_seed(42)\n",
"if torch.cuda.is_available():\n",
" torch.cuda.manual_seed_all(42)"
]
},
{
"cell_type": "code",
"execution_count": 4,
"id": "f16df21e-9797-4f78-83a1-a2943759ba55",
"metadata": {},
"outputs": [],
"source": [
"def clear_memory():\n",
" gc.collect()\n",
" torch.cuda.empty_cache()"
]
},
{
"cell_type": "code",
"execution_count": 5,
"id": "196e83da-6c8c-4cd7-bd70-2598a5e2a16a",
"metadata": {},
"outputs": [],
"source": [
"logging.basicConfig(\n",
" level=logging.INFO,\n",
" format=\"%(asctime)s - %(levelname)s - %(message)s\",\n",
")\n",
"logger = logging.getLogger(__name__)"
]
},
{
"cell_type": "code",
"execution_count": 6,
"id": "cea22b9f-f309-4151-81ac-37547c8feeb0",
"metadata": {},
"outputs": [],
"source": [
"def preprocess(text: str) -> str:\n",
" \"\"\"Remove extra whitespaces and newlines from a text string.\"\"\"\n",
" if not isinstance(text, str):\n",
" return \"\"\n",
" return re.sub(r'\\s+', ' ', text.replace('\\n', ' ')).strip()\n",
"\n",
"def clean_df(df, rename=None, drop=None, select=None):\n",
" \"\"\"\n",
" Clean and rename dataframe columns:\n",
" - drop: list of columns to drop\n",
" - rename: dict mapping old column names to new names\n",
" - select: list of columns to keep in final order\n",
" \"\"\"\n",
" if drop:\n",
" df = df.drop(columns=drop, errors='ignore')\n",
" if rename:\n",
" df = df.rename(columns=rename)\n",
" for col in ['query', 'context', 'response']:\n",
" if col in df.columns:\n",
" df[col] = df[col].apply(preprocess)\n",
" if select:\n",
" df = df[select]\n",
" return df"
]
},
{
"cell_type": "code",
"execution_count": 7,
"id": "d4eb82ce-1713-40b6-981d-43ce35aaa6f6",
"metadata": {},
"outputs": [
{
"name": "stderr",
"output_type": "stream",
"text": [
"2025-03-19 14:56:53,295 - INFO - Loading raw datasets from various sources...\n",
"2025-03-19 14:57:25,655 - INFO - Total rows before dropping duplicates: 490241\n",
"2025-03-19 14:57:27,208 - INFO - Total rows after dropping duplicates: 440785\n"
]
}
],
"source": [
"logger.info(\"Loading raw datasets from various sources...\")\n",
"\n",
"# Load datasets\n",
"df1 = pd.read_json(\"hf://datasets/Clinton/Text-to-sql-v1/texttosqlv2.jsonl\", lines=True)\n",
"df2 = pd.read_json(\"hf://datasets/b-mc2/sql-create-context/sql_create_context_v4.json\")\n",
"df3 = pd.read_parquet(\"hf://datasets/gretelai/synthetic_text_to_sql/synthetic_text_to_sql_train.snappy.parquet\")\n",
"df4 = pd.read_json(\"hf://datasets/knowrohit07/know_sql/know_sql_val3{ign}.json\")\n",
"\n",
"# Clean and rename columns to unify to 'query', 'context', 'response'\n",
"df1 = clean_df(df1, rename={'instruction': 'query', 'input': 'context'}, drop=['source', 'text'])\n",
"df2 = clean_df(df2, rename={'question': 'query', 'answer': 'response'})\n",
"df3 = clean_df(df3, rename={'sql_prompt': 'query', 'sql_context': 'context', 'sql': 'response'},\n",
" select=['query', 'context', 'response'])\n",
"df4 = clean_df(df4, rename={'question': 'query', 'answer': 'response'})\n",
"\n",
"# Concatenate all DataFrames\n",
"final_df = pd.concat([df1, df2, df3, df4], ignore_index=True)\n",
"logger.info(\"Total rows before dropping duplicates: %d\", len(final_df))\n",
"\n",
"# Force correct column order and drop rows with missing fields\n",
"final_df = final_df[['query', 'context', 'response']]\n",
"final_df = final_df.dropna(subset=['query', 'context', 'response'])\n",
"final_df = final_df.drop_duplicates()\n",
"logger.info(\"Total rows after dropping duplicates: %d\", len(final_df))"
]
},
{
"cell_type": "code",
"execution_count": 8,
"id": "8446814e-5a2c-48a4-8c01-059afcf1d3c1",
"metadata": {},
"outputs": [
{
"name": "stderr",
"output_type": "stream",
"text": [
"Token indices sequence length is longer than the specified maximum sequence length for this model (1113 > 512). Running this sequence through the model will result in indexing errors\n",
"2025-03-19 15:01:13,787 - INFO - Total rows after filtering by token length (prompt <= 500 and response <= 250 tokens): 398481\n"
]
}
],
"source": [
"tokenizer = AutoTokenizer.from_pretrained(\"google/flan-t5-base\")\n",
"\n",
"max_length_prompt = 500\n",
"max_length_response = 250\n",
"\n",
"def tokenize_length_filter(row):\n",
" start_prompt = \"Context:\\n\"\n",
" middle_prompt = \"\\n\\nQuery:\\n\"\n",
" end_prompt = \"\\n\\nResponse:\\n\"\n",
" \n",
" # Construct the prompt as used in the tokenize_function\n",
" prompt = f\"{start_prompt}{row['context']}{middle_prompt}{row['query']}{end_prompt}\"\n",
" \n",
" # Encode without truncation to get the full token count\n",
" prompt_tokens = tokenizer.encode(prompt, add_special_tokens=True, truncation=False)\n",
" response_tokens = tokenizer.encode(row['response'], add_special_tokens=True, truncation=False)\n",
" \n",
" return len(prompt_tokens) <= max_length_prompt and len(response_tokens) <= max_length_response\n",
"\n",
"final_df = final_df[final_df.apply(tokenize_length_filter, axis=1)]\n",
"logger.info(\"Total rows after filtering by token length (prompt <= %d and response <= %d tokens): %d\", \n",
" max_length_prompt, max_length_response, len(final_df))\n"
]
},
{
"cell_type": "code",
"execution_count": 9,
"id": "177e1e6d-9fbc-442d-9774-5a3e5234329f",
"metadata": {},
"outputs": [
{
"name": "stderr",
"output_type": "stream",
"text": [
"2025-03-19 15:01:13,794 - INFO - Sample from filtered final_df:\n",
" query \\\n",
"0 Name the home team for carlton away team \n",
"1 what will the population of Asia be when Latin... \n",
"2 How many faculty members do we have for each g... \n",
"\n",
" context \\\n",
"0 CREATE TABLE table_name_77 ( home_team VARCHAR... \n",
"1 CREATE TABLE table_22767 ( \"Year\" real, \"World... \n",
"2 CREATE TABLE Student ( StuID INTEGER, LName VA... \n",
"\n",
" response \n",
"0 SELECT home_team FROM table_name_77 WHERE away... \n",
"1 SELECT \"Asia\" FROM table_22767 WHERE \"Latin Am... \n",
"2 SELECT Sex, COUNT(*) FROM Faculty GROUP BY Sex... \n"
]
}
],
"source": [
"logger.info(\"Sample from filtered final_df:\\n%s\", final_df.head(3))\n",
"clear_memory()"
]
},
{
"cell_type": "code",
"execution_count": 10,
"id": "0b639efe-ebeb-4b34-bc3f-accf776ba0da",
"metadata": {},
"outputs": [
{
"name": "stderr",
"output_type": "stream",
"text": [
"2025-03-19 15:01:14,006 - INFO - Final split sizes: Train: 338708, Test: 39848, Validation: 19925\n"
]
},
{
"data": {
"application/vnd.jupyter.widget-view+json": {
"model_id": "81e753f720e44f40b5f0dfa5263e2bf5",
"version_major": 2,
"version_minor": 0
},
"text/plain": [
"Saving the dataset (0/1 shards): 0%| | 0/338708 [00:00<?, ? examples/s]"
]
},
"metadata": {},
"output_type": "display_data"
},
{
"data": {
"application/vnd.jupyter.widget-view+json": {
"model_id": "59b1ce0d9ee548668dbc87b99d6e0951",
"version_major": 2,
"version_minor": 0
},
"text/plain": [
"Saving the dataset (0/1 shards): 0%| | 0/39848 [00:00<?, ? examples/s]"
]
},
"metadata": {},
"output_type": "display_data"
},
{
"data": {
"application/vnd.jupyter.widget-view+json": {
"model_id": "4a378405a0a24c13a81fc853550d01d6",
"version_major": 2,
"version_minor": 0
},
"text/plain": [
"Saving the dataset (0/1 shards): 0%| | 0/19925 [00:00<?, ? examples/s]"
]
},
"metadata": {},
"output_type": "display_data"
},
{
"name": "stderr",
"output_type": "stream",
"text": [
"2025-03-19 15:01:15,490 - INFO - Merged and Saved Dataset Successfully!\n",
"2025-03-19 15:01:15,497 - INFO - Dataset summary: DatasetDict({\n",
" train: Dataset({\n",
" features: ['query', 'context', 'response'],\n",
" num_rows: 338708\n",
" })\n",
" test: Dataset({\n",
" features: ['query', 'context', 'response'],\n",
" num_rows: 39848\n",
" })\n",
" validation: Dataset({\n",
" features: ['query', 'context', 'response'],\n",
" num_rows: 19925\n",
" })\n",
"})\n"
]
}
],
"source": [
"def split_dataframe(df, train_frac=0.85, test_frac=0.1, val_frac=0.05):\n",
" n = len(df)\n",
" train_end = int(n * train_frac)\n",
" test_end = train_end + int(n * test_frac)\n",
" train_df = df.iloc[:train_end].reset_index(drop=True)\n",
" test_df = df.iloc[train_end:test_end].reset_index(drop=True)\n",
" val_df = df.iloc[test_end:].reset_index(drop=True)\n",
" return train_df, test_df, val_df\n",
"\n",
"train_df, test_df, val_df = split_dataframe(final_df)\n",
"logger.info(\"Final split sizes: Train: %d, Test: %d, Validation: %d\", len(train_df), len(test_df), len(val_df))\n",
"\n",
"# Convert splits to Hugging Face Datasets\n",
"train_dataset = Dataset.from_pandas(train_df)\n",
"test_dataset = Dataset.from_pandas(test_df)\n",
"val_dataset = Dataset.from_pandas(val_df)\n",
"\n",
"dataset = DatasetDict({\n",
" 'train': train_dataset,\n",
" 'test': test_dataset,\n",
" 'validation': val_dataset\n",
"})\n",
"\n",
"dataset.save_to_disk(\"merged_dataset\")\n",
"logger.info(\"Merged and Saved Dataset Successfully!\")\n",
"logger.info(\"Dataset summary: %s\", dataset)\n",
"clear_memory()"
]
},
{
"cell_type": "code",
"execution_count": 11,
"id": "9f6e1095-d72d-4e22-b20d-683f1f84544c",
"metadata": {},
"outputs": [
{
"name": "stderr",
"output_type": "stream",
"text": [
"2025-03-19 15:01:15,843 - INFO - Reloaded dataset from disk. Example from test split:\n",
"{'query': \"Show the name and type of military cyber commands in the 'Military_Cyber_Commands' table.\", 'context': \"CREATE SCHEMA IF NOT EXISTS defense_security;CREATE TABLE IF NOT EXISTS defense_security.Military_Cyber_Commands (id INT PRIMARY KEY, command_name VARCHAR(255), type VARCHAR(255));INSERT INTO defense_security.Military_Cyber_Commands (id, command_name, type) VALUES (1, 'USCYBERCOM', 'Defensive Cyber Operations'), (2, 'JTF-CND', 'Offensive Cyber Operations'), (3, '10th Fleet', 'Network Warfare');\", 'response': 'SELECT command_name, type FROM defense_security.Military_Cyber_Commands;'}\n",
"2025-03-19 15:01:16,155 - INFO - Loaded Tokenized Dataset from disk.\n",
"2025-03-19 15:01:16,159 - INFO - Final tokenized dataset splits: dict_keys(['train', 'test', 'validation'])\n",
"2025-03-19 15:01:16,167 - INFO - Sample tokenized record from train split:\n",
"{'input_ids': tensor([ 1193, 6327, 10, 205, 4386, 6048, 332, 17098, 953, 834,\n",
" 4350, 834, 4013, 41, 234, 834, 11650, 584, 4280, 28027,\n",
" 6, 550, 834, 11650, 584, 4280, 28027, 3, 61, 3,\n",
" 27569, 10, 5570, 8, 234, 372, 21, 443, 7377, 550,\n",
" 372, 16361, 10, 3, 1, 0, 0, 0, 0, 0,\n",
" 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,\n",
" 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,\n",
" 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,\n",
" 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,\n",
" 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,\n",
" 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,\n",
" 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,\n",
" 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,\n",
" 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,\n",
" 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,\n",
" 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,\n",
" 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,\n",
" 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,\n",
" 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,\n",
" 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,\n",
" 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,\n",
" 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,\n",
" 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,\n",
" 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,\n",
" 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,\n",
" 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,\n",
" 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,\n",
" 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,\n",
" 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,\n",
" 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,\n",
" 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,\n",
" 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,\n",
" 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,\n",
" 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,\n",
" 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,\n",
" 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,\n",
" 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,\n",
" 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,\n",
" 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,\n",
" 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,\n",
" 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,\n",
" 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,\n",
" 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,\n",
" 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,\n",
" 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,\n",
" 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,\n",
" 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,\n",
" 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,\n",
" 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,\n",
" 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,\n",
" 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,\n",
" 0, 0]), 'attention_mask': tensor([1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1,\n",
" 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 0, 0, 0,\n",
" 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,\n",
" 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,\n",
" 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,\n",
" 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,\n",
" 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,\n",
" 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,\n",
" 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,\n",
" 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,\n",
" 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,\n",
" 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,\n",
" 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,\n",
" 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,\n",
" 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,\n",
" 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,\n",
" 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,\n",
" 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,\n",
" 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,\n",
" 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,\n",
" 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,\n",
" 0, 0, 0, 0, 0, 0, 0, 0]), 'labels': tensor([ 3, 23143, 14196, 234, 834, 11650, 21680, 953, 834, 4350,\n",
" 834, 4013, 549, 17444, 427, 550, 834, 11650, 3274, 96,\n",
" 1720, 7377, 121, 1, -100, -100, -100, -100, -100, -100,\n",
" -100, -100, -100, -100, -100, -100, -100, -100, -100, -100,\n",
" -100, -100, -100, -100, -100, -100, -100, -100, -100, -100,\n",
" -100, -100, -100, -100, -100, -100, -100, -100, -100, -100,\n",
" -100, -100, -100, -100, -100, -100, -100, -100, -100, -100,\n",
" -100, -100, -100, -100, -100, -100, -100, -100, -100, -100,\n",
" -100, -100, -100, -100, -100, -100, -100, -100, -100, -100,\n",
" -100, -100, -100, -100, -100, -100, -100, -100, -100, -100,\n",
" -100, -100, -100, -100, -100, -100, -100, -100, -100, -100,\n",
" -100, -100, -100, -100, -100, -100, -100, -100, -100, -100,\n",
" -100, -100, -100, -100, -100, -100, -100, -100, -100, -100,\n",
" -100, -100, -100, -100, -100, -100, -100, -100, -100, -100,\n",
" -100, -100, -100, -100, -100, -100, -100, -100, -100, -100,\n",
" -100, -100, -100, -100, -100, -100, -100, -100, -100, -100,\n",
" -100, -100, -100, -100, -100, -100, -100, -100, -100, -100,\n",
" -100, -100, -100, -100, -100, -100, -100, -100, -100, -100,\n",
" -100, -100, -100, -100, -100, -100, -100, -100, -100, -100,\n",
" -100, -100, -100, -100, -100, -100, -100, -100, -100, -100,\n",
" -100, -100, -100, -100, -100, -100, -100, -100, -100, -100,\n",
" -100, -100, -100, -100, -100, -100, -100, -100, -100, -100,\n",
" -100, -100, -100, -100, -100, -100, -100, -100, -100, -100,\n",
" -100, -100, -100, -100, -100, -100, -100, -100, -100, -100,\n",
" -100, -100, -100, -100, -100, -100, -100, -100, -100, -100,\n",
" -100, -100, -100, -100, -100, -100])}\n"
]
}
],
"source": [
"dataset = load_from_disk(\"merged_dataset\")\n",
"logger.info(\"Reloaded dataset from disk. Example from test split:\\n%s\", dataset['test'][0])\n",
"\n",
"model_name = \"google/flan-t5-base\"\n",
"tokenizer = AutoTokenizer.from_pretrained(model_name)\n",
"\n",
"def tokenize_function(batch: dict) -> dict:\n",
" \"\"\"\n",
" Tokenizes a batch of examples for T5 fine-tuning.\n",
" Constructs a prompt in the format:\n",
" Context:\n",
" <context>\n",
" \n",
" Query:\n",
" <query>\n",
" \n",
" Response:\n",
" \"\"\"\n",
" start_prompt = \"Context:\\n\"\n",
" middle_prompt = \"\\n\\nQuery:\\n\"\n",
" end_prompt = \"\\n\\nResponse:\\n\"\n",
"\n",
" prompts = [\n",
" f\"{start_prompt}{ctx}{middle_prompt}{qry}{end_prompt}\"\n",
" for ctx, qry in zip(batch['context'], batch['query'])\n",
" ]\n",
"\n",
" tokenized_inputs = tokenizer(\n",
" prompts,\n",
" padding=\"max_length\",\n",
" truncation=True,\n",
" max_length=512\n",
" )\n",
" tokenized_labels = tokenizer(\n",
" batch['response'],\n",
" padding=\"max_length\",\n",
" truncation=True,\n",
" max_length=256\n",
" )\n",
" labels = [\n",
" [-100 if token == tokenizer.pad_token_id else token for token in seq]\n",
" for seq in tokenized_labels['input_ids']\n",
" ]\n",
"\n",
" batch['input_ids'] = tokenized_inputs['input_ids']\n",
" batch['attention_mask'] = tokenized_inputs['attention_mask']\n",
" batch['labels'] = labels\n",
" return batch\n",
"\n",
"try:\n",
" tokenized_datasets = load_from_disk(\"tokenized_datasets\")\n",
" logger.info(\"Loaded Tokenized Dataset from disk.\")\n",
"except Exception as e:\n",
" logger.info(\"Tokenized dataset not found. Creating a new one...\")\n",
" tokenized_datasets = dataset.map(\n",
" tokenize_function,\n",
" batched=True,\n",
" remove_columns=['query', 'context', 'response'],\n",
" num_proc=8\n",
" )\n",
" tokenized_datasets.save_to_disk(\"tokenized_datasets\")\n",
" logger.info(\"Tokenized and Saved Dataset.\")\n",
"\n",
"tokenized_datasets.set_format(\"torch\")\n",
"\n",
"logger.info(\"Final tokenized dataset splits: %s\", tokenized_datasets.keys())\n",
"logger.info(\"Sample tokenized record from train split:\\n%s\", tokenized_datasets['train'][0])"
]
},
{
"cell_type": "code",
"execution_count": 12,
"id": "7f004e55-181c-47aa-9f3e-c7c1ceae780c",
"metadata": {},
"outputs": [
{
"name": "stdout",
"output_type": "stream",
"text": [
"----------------------------------------------------------------------------------------------------\n",
"INPUT PROMPT:\n",
"Context:\n",
"CREATE SCHEMA IF NOT EXISTS defense_security;CREATE TABLE IF NOT EXISTS defense_security.Military_Cyber_Commands (id INT PRIMARY KEY, command_name VARCHAR(255), type VARCHAR(255));INSERT INTO defense_security.Military_Cyber_Commands (id, command_name, type) VALUES (1, 'USCYBERCOM', 'Defensive Cyber Operations'), (2, 'JTF-CND', 'Offensive Cyber Operations'), (3, '10th Fleet', 'Network Warfare');\n",
"\n",
"Query:\n",
"Show the name and type of military cyber commands in the 'Military_Cyber_Commands' table.\n",
"\n",
"Response:\n",
"\n",
"----------------------------------------------------------------------------------------------------\n",
"BASELINE HUMAN ANSWER:\n",
"SELECT command_name, type FROM defense_security.Military_Cyber_Commands;\n",
"\n",
"----------------------------------------------------------------------------------------------------\n",
"MODEL GENERATION - ZERO SHOT:\n",
"USCYBERCOM, JTF-CND, Offensive Cyber Operations, 10th Fleet, Network Warfare\n"
]
}
],
"source": [
"model_name = 'google/flan-t5-base'\n",
"tokenizer = AutoTokenizer.from_pretrained(model_name)\n",
"original_model = AutoModelForSeq2SeqLM.from_pretrained(model_name, torch_dtype=torch.bfloat16)\n",
"original_model = original_model.to(device)\n",
"\n",
"index = 0\n",
"query = dataset['test'][index]['query']\n",
"context = dataset['test'][index]['context']\n",
"response = dataset['test'][index]['response']\n",
"\n",
"prompt = f\"\"\"Context:\n",
"{context}\n",
"\n",
"Query:\n",
"{query}\n",
"\n",
"Response:\n",
"\"\"\"\n",
"inputs = tokenizer(prompt, return_tensors='pt').to(device)\n",
"baseline_output = tokenizer.decode(\n",
" original_model.generate(\n",
" inputs[\"input_ids\"],\n",
" max_new_tokens=200,\n",
" )[0],\n",
" skip_special_tokens=True\n",
")\n",
"dash_line = '-' * 100\n",
"print(dash_line)\n",
"print(f'INPUT PROMPT:\\n{prompt}')\n",
"print(dash_line)\n",
"print(f'BASELINE HUMAN ANSWER:\\n{response}\\n')\n",
"print(dash_line)\n",
"print(f'MODEL GENERATION - ZERO SHOT:\\n{baseline_output}')\n",
"clear_memory()"
]
},
{
"cell_type": "code",
"execution_count": 13,
"id": "f50e56c7-98b3-42bc-9129-89f3eff802e7",
"metadata": {},
"outputs": [
{
"name": "stderr",
"output_type": "stream",
"text": [
"2025-03-19 15:01:30,827 - INFO - Attempting to load the fine-tuned model...\n",
"2025-03-19 15:01:32,195 - INFO - Fine-tuned model loaded successfully.\n"
]
}
],
"source": [
"import math\n",
"\n",
"try:\n",
" logger.info(\"Attempting to load the fine-tuned model...\")\n",
" finetuned_model = AutoModelForSeq2SeqLM.from_pretrained(\"text2sql_flant5base_finetuned\")\n",
" tokenizer = AutoTokenizer.from_pretrained(\"google/flan-t5-base\")\n",
" finetuned_model = finetuned_model.to(device)\n",
" to_train = False\n",
" logger.info(\"Fine-tuned model loaded successfully.\")\n",
"except Exception as e:\n",
" logger.info(\"Fine-tuned model not found.\")\n",
" logger.info(\"Initializing model and tokenizer for QLORA fine-tuning...\")\n",
" to_train = True\n",
"\n",
" quant_config = BitsAndBytesConfig(\n",
" load_in_4bit=True,\n",
" bnb_4bit_quant_type=\"nf4\",\n",
" bnb_4bit_use_double_quant=True,\n",
" bnb_4bit_compute_dtype=torch.bfloat16,\n",
" )\n",
"\n",
" finetuned_model = AutoModelForSeq2SeqLM.from_pretrained(\n",
" model_name,\n",
" quantization_config=quant_config,\n",
" device_map=\"auto\",\n",
" torch_dtype=torch.bfloat16,\n",
" )\n",
" finetuned_model = prepare_model_for_kbit_training(finetuned_model)\n",
" \n",
" lora_config = LoraConfig(\n",
" r=32,\n",
" lora_alpha=64,\n",
" target_modules=[\"q\", \"v\"],\n",
" lora_dropout=0.1,\n",
" bias=\"none\",\n",
" task_type=\"SEQ_2_SEQ_LM\"\n",
" )\n",
" finetuned_model = get_peft_model(finetuned_model, lora_config)\n",
" tokenizer = AutoTokenizer.from_pretrained(model_name)\n",
" logger.info(\"Base model loaded and prepared for QLORA fine-tuning.\")\n",
" clear_memory()\n",
"\n",
"if to_train:\n",
" output_dir = f\"./sql-training-{int(time.time())}\"\n",
" logger.info(\"Starting training. Output directory: %s\", output_dir)\n",
"\n",
" # Compute total training steps:\n",
" num_train_samples = len(tokenized_datasets[\"train\"])\n",
" per_device_train_batch_size = 64\n",
" per_device_eval_batch_size = 64\n",
" num_train_epochs = 6\n",
" # Assuming no gradient accumulation beyond the per-device batch size\n",
" total_steps = math.ceil(num_train_samples / per_device_train_batch_size) * num_train_epochs\n",
" # Set warmup steps as 10% of total steps (adjust as needed)\n",
" warmup_steps = int(total_steps * 0.1)\n",
" \n",
" logger.info(\"Total training steps: %d, Warmup steps (10%%): %d\", total_steps, warmup_steps)\n",
" \n",
" training_args = TrainingArguments(\n",
" output_dir=output_dir,\n",
" gradient_checkpointing=True,\n",
" gradient_checkpointing_kwargs={\"use_reentrant\": True},\n",
" gradient_accumulation_steps = 2,\n",
" learning_rate=2e-4,\n",
" optim=\"adamw_bnb_8bit\", # Memory-efficient optimizer\n",
" num_train_epochs=num_train_epochs,\n",
" per_device_train_batch_size=per_device_train_batch_size,\n",
" per_device_eval_batch_size=per_device_eval_batch_size,\n",
" weight_decay=0.01,\n",
" logging_steps=200, \n",
" logging_dir=f\"{output_dir}/logs\",\n",
" eval_strategy=\"epoch\", # Evaluate at the end of each epoch\n",
" save_strategy=\"epoch\", # Save the model at the end of each epoch\n",
" save_total_limit=3,\n",
" load_best_model_at_end=True,\n",
" metric_for_best_model=\"eval_loss\",\n",
" bf16=True, \n",
" warmup_ratio=0.1, # Warmup 10% of total steps\n",
" lr_scheduler_type=\"cosine\",\n",
" )\n",
" trainer = Trainer(\n",
" model=finetuned_model,\n",
" args=training_args,\n",
" train_dataset=tokenized_datasets[\"train\"],\n",
" eval_dataset=tokenized_datasets[\"validation\"],\n",
" callbacks=[EarlyStoppingCallback(early_stopping_patience=2)],\n",
" )\n",
" logger.info(\"Beginning fine-tuning...\")\n",
" trainer.train()\n",
" logger.info(\"Training completed.\")\n",
" save_path = \"text2sql_flant5base_finetuned\"\n",
" finetuned_model.save_pretrained(save_path)\n",
" logger.info(\"Model saved to %s\", save_path)\n",
" clear_memory()"
]
},
{
"cell_type": "code",
"execution_count": 14,
"id": "f364eb6b-56cb-4533-8ef6-b5e7f56895aa",
"metadata": {},
"outputs": [
{
"name": "stderr",
"output_type": "stream",
"text": [
"2025-03-19 15:01:32,235 - INFO - Running inference on 5 examples (displaying real responses).\n",
"/venv/main/lib/python3.10/site-packages/transformers/generation/configuration_utils.py:629: UserWarning: `do_sample` is set to `False`. However, `temperature` is set to `0.1` -- this flag is only used in sample-based generation modes. You should set `do_sample=True` or unset `temperature`.\n",
" warnings.warn(\n"
]
},
{
"name": "stdout",
"output_type": "stream",
"text": [
"\n",
"====================================================================================================\n",
"----------------------------------------------------------------------------------------------------\n",
"Example 1\n",
"----------------------------------------------------------------------------------------------------\n",
"INPUT PROMPT:\n",
"Context:\n",
"CREATE SCHEMA IF NOT EXISTS defense_security;CREATE TABLE IF NOT EXISTS defense_security.Military_Cyber_Commands (id INT PRIMARY KEY, command_name VARCHAR(255), type VARCHAR(255));INSERT INTO defense_security.Military_Cyber_Commands (id, command_name, type) VALUES (1, 'USCYBERCOM', 'Defensive Cyber Operations'), (2, 'JTF-CND', 'Offensive Cyber Operations'), (3, '10th Fleet', 'Network Warfare');\n",
"\n",
"Query:\n",
"Show the name and type of military cyber commands in the 'Military_Cyber_Commands' table.\n",
"\n",
"Response:\n",
"\n",
"----------------------------------------------------------------------------------------------------\n",
"HUMAN RESPONSE:\n",
"SELECT command_name, type FROM defense_security.Military_Cyber_Commands;\n",
"----------------------------------------------------------------------------------------------------\n",
"ORIGINAL MODEL OUTPUT:\n",
"USCYBERCOM, JTF-CND, Offensive Cyber Operations\n",
"----------------------------------------------------------------------------------------------------\n",
"FINE-TUNED MODEL OUTPUT:\n",
"SELECT command_name, type FROM defense_security.Military_Cyber_Commands;\n",
"====================================================================================================\n",
"\n",
"----------------------------------------------------------------------------------------------------\n",
"Example 2\n",
"----------------------------------------------------------------------------------------------------\n",
"INPUT PROMPT:\n",
"Context:\n",
"CREATE TABLE incidents (id INT, cause VARCHAR(255), cost INT, date DATE); INSERT INTO incidents (id, cause, cost, date) VALUES (1, 'insider threat', 10000, '2022-01-01'); INSERT INTO incidents (id, cause, cost, date) VALUES (2, 'phishing', 5000, '2022-01-02');\n",
"\n",
"Query:\n",
"Find the total cost of all security incidents caused by insider threats in the last 6 months\n",
"\n",
"Response:\n",
"\n",
"----------------------------------------------------------------------------------------------------\n",
"HUMAN RESPONSE:\n",
"SELECT SUM(cost) FROM incidents WHERE cause = 'insider threat' AND date >= DATE_SUB(CURRENT_DATE, INTERVAL 6 MONTH);\n",
"----------------------------------------------------------------------------------------------------\n",
"ORIGINAL MODEL OUTPUT:\n",
"10000, 2022-01-01\n",
"----------------------------------------------------------------------------------------------------\n",
"FINE-TUNED MODEL OUTPUT:\n",
"SELECT SUM(cost) FROM incidents WHERE cause = 'insider threat' AND date >= DATE_SUB(CURRENT_DATE, INTERVAL 6 MONTH);\n",
"====================================================================================================\n",
"\n",
"----------------------------------------------------------------------------------------------------\n",
"Example 3\n",
"----------------------------------------------------------------------------------------------------\n",
"INPUT PROMPT:\n",
"Context:\n",
"CREATE TABLE libraries (name VARCHAR(255), state VARCHAR(255), population DECIMAL(10,2), libraries DECIMAL(5,2)); INSERT INTO libraries (name, state, population, libraries) VALUES ('Library1', 'California', 39512223, 3154), ('Library2', 'Texas', 29528404, 2212), ('Library3', 'Florida', 21644287, 1835);\n",
"\n",
"Query:\n",
"Show the top 3 states with the most public libraries per capita.\n",
"\n",
"Response:\n",
"\n",
"----------------------------------------------------------------------------------------------------\n",
"HUMAN RESPONSE:\n",
"SELECT state, (libraries / population) AS libraries_per_capita FROM libraries ORDER BY libraries_per_capita DESC LIMIT 3;\n",
"----------------------------------------------------------------------------------------------------\n",
"ORIGINAL MODEL OUTPUT:\n",
"California, 39512223, 3154\n",
"----------------------------------------------------------------------------------------------------\n",
"FINE-TUNED MODEL OUTPUT:\n",
"SELECT state, population, RANK() OVER (ORDER BY population DESC) as rank FROM libraries GROUP BY state ORDER BY rank DESC LIMIT 3;\n",
"====================================================================================================\n",
"\n",
"----------------------------------------------------------------------------------------------------\n",
"Example 4\n",
"----------------------------------------------------------------------------------------------------\n",
"INPUT PROMPT:\n",
"Context:\n",
"CREATE TABLE users (id INT, location VARCHAR(50)); CREATE TABLE posts (id INT, user_id INT, created_at DATETIME);\n",
"\n",
"Query:\n",
"What is the total number of posts made by users located in Australia, in the last month?\n",
"\n",
"Response:\n",
"\n",
"----------------------------------------------------------------------------------------------------\n",
"HUMAN RESPONSE:\n",
"SELECT COUNT(posts.id) FROM posts INNER JOIN users ON posts.user_id = users.id WHERE users.location = 'Australia' AND posts.created_at >= DATE_SUB(NOW(), INTERVAL 1 MONTH);\n",
"----------------------------------------------------------------------------------------------------\n",
"ORIGINAL MODEL OUTPUT:\n",
"The total number of posts made by users located in Australia is 50.\n",
"----------------------------------------------------------------------------------------------------\n",
"FINE-TUNED MODEL OUTPUT:\n",
"SELECT COUNT(*) FROM posts p JOIN users u ON p.user_id = u.id WHERE u.location = 'Australia' AND p.created_at >= DATE_SUB(CURRENT_DATE, INTERVAL 1 MONTH);\n",
"====================================================================================================\n",
"\n"
]
},
{
"name": "stderr",
"output_type": "stream",
"text": [
"2025-03-19 15:01:40,448 - INFO - Starting evaluation on the full test set using batching.\n"
]
},
{
"name": "stdout",
"output_type": "stream",
"text": [
"----------------------------------------------------------------------------------------------------\n",
"Example 5\n",
"----------------------------------------------------------------------------------------------------\n",
"INPUT PROMPT:\n",
"Context:\n",
"CREATE TABLE WindFarms (FarmID INT, FarmName VARCHAR(255), Capacity DECIMAL(5,2), Country VARCHAR(255)); INSERT INTO WindFarms (FarmID, FarmName, Capacity, Country) VALUES (1, 'WindFarm1', 150, 'USA'), (2, 'WindFarm2', 200, 'Canada'), (3, 'WindFarm3', 120, 'Mexico');\n",
"\n",
"Query:\n",
"List the total installed capacity of wind farms in the WindEnergy schema for each country?\n",
"\n",
"Response:\n",
"\n",
"----------------------------------------------------------------------------------------------------\n",
"HUMAN RESPONSE:\n",
"SELECT Country, SUM(Capacity) as TotalCapacity FROM WindFarms GROUP BY Country;\n",
"----------------------------------------------------------------------------------------------------\n",
"ORIGINAL MODEL OUTPUT:\n",
"1, 150, USA, 2, 200, Canada, 3, 120, Mexico\n",
"----------------------------------------------------------------------------------------------------\n",
"FINE-TUNED MODEL OUTPUT:\n",
"SELECT Country, SUM(Capacity) FROM WindFarms GROUP BY Country;\n",
"====================================================================================================\n",
"\n"
]
},
{
"data": {
"application/vnd.jupyter.widget-view+json": {
"model_id": "a7beecee09a34f9790be1e4538a87442",
"version_major": 2,
"version_minor": 0
},
"text/plain": [
"Downloading builder script: 0%| | 0.00/5.94k [00:00<?, ?B/s]"
]
},
"metadata": {},
"output_type": "display_data"
},
{
"data": {
"application/vnd.jupyter.widget-view+json": {
"model_id": "763373c451c94f5e92bc6a6253109275",
"version_major": 2,
"version_minor": 0
},
"text/plain": [
"Downloading extra modules: 0%| | 0.00/1.55k [00:00<?, ?B/s]"
]
},
"metadata": {},
"output_type": "display_data"
},
{
"data": {
"application/vnd.jupyter.widget-view+json": {
"model_id": "afdce82cb8964da788756d783539ee8d",
"version_major": 2,
"version_minor": 0
},
"text/plain": [
"Downloading extra modules: 0%| | 0.00/3.34k [00:00<?, ?B/s]"
]
},
"metadata": {},
"output_type": "display_data"
},
{
"name": "stderr",
"output_type": "stream",
"text": [
"2025-03-19 16:47:58,173 - INFO - Using default tokenizer.\n",
"2025-03-19 16:49:07,668 - INFO - Using default tokenizer.\n"
]
},
{
"name": "stdout",
"output_type": "stream",
"text": [
"\n",
"====================================================================================================\n",
"Evaluation Metrics:\n",
"====================================================================================================\n",
"ORIGINAL MODEL:\n",
" ROUGE: {'rouge1': np.float64(0.05646642898660111), 'rouge2': np.float64(0.01562815013068162), 'rougeL': np.float64(0.05031267225420556), 'rougeLsum': np.float64(0.05036072587316542)}\n",
" BLEU: {'bleu': 0.003142147128241449, 'precisions': [0.12293406776920406, 0.03289697910893642, 0.018512080104175887, 0.008342750223825794], 'brevity_penalty': 0.11177079327444009, 'length_ratio': 0.3133514352662089, 'translation_length': 377251, 'reference_length': 1203923}\n",
" Fuzzy Match Score: 13.98%\n",
" Exact Match Accuracy: 0.00%\n",
"\n",
"FINE-TUNED MODEL:\n",
" ROUGE: {'rouge1': np.float64(0.7538800834024002), 'rouge2': np.float64(0.6103863808522726), 'rougeL': np.float64(0.7262841884754194), 'rougeLsum': np.float64(0.7261852209847466)}\n",
" BLEU: {'bleu': 0.4719774431701209, 'precisions': [0.7603153442288385, 0.598309257795389, 0.5021259810303533, 0.42128998564638875], 'brevity_penalty': 0.8474086962179814, 'length_ratio': 0.8579477258927689, 'translation_length': 1032903, 'reference_length': 1203923}\n",
" Fuzzy Match Score: 85.62%\n",
" Exact Match Accuracy: 18.29%\n",
"====================================================================================================\n"
]
}
],
"source": [
"import logging\n",
"import re\n",
"import pandas as pd\n",
"from rapidfuzz import fuzz\n",
"import evaluate\n",
"\n",
"# Assuming tokenizer, device, original_model, finetuned_model, and dataset are already defined.\n",
"# Define a helper function for output post-processing.\n",
"def post_process_output(output_text: str) -> str:\n",
" \"\"\"Post-process the generated output to remove repeated text.\"\"\"\n",
" # Keep only the first valid SQL query (everything before the first semicolon)\n",
" return output_text.split(\";\")[0] + \";\" if \";\" in output_text else output_text\n",
"\n",
"# Define a helper function for generating outputs with the given generation parameters.\n",
"def generate_with_params(model, input_ids):\n",
" generated_ids = model.generate(\n",
" input_ids=input_ids,\n",
" max_new_tokens=100, \n",
" num_beams=5,\n",
" repetition_penalty=1.2,\n",
" temperature=0.1,\n",
" early_stopping=True\n",
" )\n",
" # Decode and post-process output\n",
" output_text = tokenizer.decode(generated_ids[0], skip_special_tokens=True)\n",
" return output_text\n",
"\n",
"# Helper functions for SQL normalization and evaluation metrics\n",
"def normalize_sql(sql):\n",
" \"\"\"Normalize SQL by stripping whitespace and lowercasing.\"\"\"\n",
" return \" \".join(sql.strip().lower().split())\n",
"\n",
"def compute_exact_match(predictions, references):\n",
" \"\"\"Computes the exact match accuracy after normalization.\"\"\"\n",
" matches = sum(1 for pred, ref in zip(predictions, references)\n",
" if normalize_sql(pred) == normalize_sql(ref))\n",
" return (matches / len(predictions)) * 100 if predictions else 0\n",
"\n",
"def compute_fuzzy_match(predictions, references):\n",
" \"\"\"Computes a soft matching score using token_set_ratio from rapidfuzz.\"\"\"\n",
" scores = [fuzz.token_set_ratio(pred, ref) for pred, ref in zip(predictions, references)]\n",
" return sum(scores) / len(scores) if scores else 0\n",
"\n",
"# Dummy function to free up memory if needed.\n",
"def clear_memory():\n",
" # If using torch.cuda, you can clear cache:\n",
" # torch.cuda.empty_cache()\n",
" pass\n",
"\n",
"logger = logging.getLogger(__name__)\n",
"logger.setLevel(logging.INFO)\n",
"\n",
"# --- Part A: Inference on 5 Examples with Real Responses ---\n",
"logger.info(\"Running inference on 5 examples (displaying real responses).\")\n",
"\n",
"num_examples = 5\n",
"sample_queries = dataset[\"test\"][:num_examples][\"query\"]\n",
"sample_contexts = dataset[\"test\"][:num_examples][\"context\"]\n",
"sample_human_responses = dataset[\"test\"][:num_examples][\"response\"]\n",
"\n",
"print(\"\\n\" + \"=\" * 100)\n",
"for idx in range(num_examples):\n",
" prompt = f\"\"\"Context:\n",
"{sample_contexts[idx]}\n",
"\n",
"Query:\n",
"{sample_queries[idx]}\n",
"\n",
"Response:\n",
"\"\"\"\n",
" # Tokenize the prompt and move to device\n",
" inputs = tokenizer(prompt, return_tensors=\"pt\", truncation=True, max_length=512).to(device)\n",
" \n",
" # Generate outputs using the modified generation parameters\n",
" orig_out = generate_with_params(original_model, inputs[\"input_ids\"])\n",
" finetuned_out = post_process_output(generate_with_params(finetuned_model, inputs[\"input_ids\"]))\n",
" \n",
" print(\"-\" * 100)\n",
" print(f\"Example {idx+1}\")\n",
" print(\"-\" * 100)\n",
" print(\"INPUT PROMPT:\")\n",
" print(prompt)\n",
" print(\"-\" * 100)\n",
" print(\"HUMAN RESPONSE:\")\n",
" print(sample_human_responses[idx])\n",
" print(\"-\" * 100)\n",
" print(\"ORIGINAL MODEL OUTPUT:\")\n",
" print(orig_out)\n",
" print(\"-\" * 100)\n",
" print(\"FINE-TUNED MODEL OUTPUT:\")\n",
" print(finetuned_out)\n",
" print(\"=\" * 100 + \"\\n\")\n",
" clear_memory()\n",
"\n",
"# --- Part B: Evaluation on Full Test Set with Batching (Optimized) ---\n",
"logger.info(\"Starting evaluation on the full test set using batching.\")\n",
"\n",
"all_human_responses = []\n",
"all_original_responses = []\n",
"all_finetuned_responses = []\n",
"\n",
"batch_size = 128 # Adjust based on GPU memory\n",
"test_dataset = dataset[\"test\"]\n",
"\n",
"for i in range(0, len(test_dataset), batch_size):\n",
" # Slicing the dataset returns a dict of lists\n",
" batch = test_dataset[i:i + batch_size]\n",
" \n",
" # Construct prompts for each example in the batch\n",
" prompts = [\n",
" f\"Context:\\n{batch['context'][j]}\\n\\nQuery:\\n{batch['query'][j]}\\n\\nResponse:\"\n",
" for j in range(len(batch[\"context\"]))\n",
" ]\n",
" \n",
" # Extend human responses\n",
" all_human_responses.extend(batch[\"response\"])\n",
" \n",
" # Tokenize the batch of prompts with padding and truncation\n",
" inputs = tokenizer(prompts, return_tensors=\"pt\", padding=True, truncation=True, max_length=512).to(device)\n",
" \n",
" # Generate outputs for the batch for both models\n",
" orig_ids = original_model.generate(\n",
" input_ids=inputs[\"input_ids\"],\n",
" max_new_tokens=100,\n",
" num_beams=5,\n",
" repetition_penalty=1.2,\n",
" temperature=0.1,\n",
" early_stopping=True\n",
" )\n",
" finetuned_ids = finetuned_model.generate(\n",
" input_ids=inputs[\"input_ids\"],\n",
" max_new_tokens=100,\n",
" num_beams=5,\n",
" repetition_penalty=1.2,\n",
" temperature=0.1,\n",
" early_stopping=True\n",
" )\n",
" \n",
" # Decode and post-process each sample in the batch\n",
" orig_texts = [tokenizer.decode(ids, skip_special_tokens=True) for ids in orig_ids]\n",
" finetuned_texts = [post_process_output(tokenizer.decode(ids, skip_special_tokens=True)) for ids in finetuned_ids]\n",
" \n",
" all_original_responses.extend(orig_texts)\n",
" all_finetuned_responses.extend(finetuned_texts)\n",
" clear_memory()\n",
"\n",
"# Create a DataFrame for a quick comparison of results\n",
"zipped_all = list(zip(all_human_responses, all_original_responses, all_finetuned_responses))\n",
"df_full = pd.DataFrame(zipped_all, columns=[\"Human Response\", \"Original Model Output\", \"Fine-Tuned Model Output\"])\n",
"df_full.to_csv('evaluation_results.csv', index=False)\n",
"clear_memory()\n",
"\n",
"# --- Compute Evaluation Metrics ---\n",
"rouge = evaluate.load(\"rouge\")\n",
"bleu = evaluate.load(\"bleu\")\n",
"\n",
"# Compute metrics for the original (non-fine-tuned) model\n",
"orig_rouge = rouge.compute(\n",
" predictions=all_original_responses,\n",
" references=all_human_responses,\n",
" use_aggregator=True,\n",
" use_stemmer=True,\n",
")\n",
"orig_bleu = bleu.compute(\n",
" predictions=all_original_responses,\n",
" references=[[ref] for ref in all_human_responses]\n",
")\n",
"orig_fuzzy = compute_fuzzy_match(all_original_responses, all_human_responses)\n",
"orig_exact = compute_exact_match(all_original_responses, all_human_responses)\n",
"\n",
"# Compute metrics for the fine-tuned model\n",
"finetuned_rouge = rouge.compute(\n",
" predictions=all_finetuned_responses,\n",
" references=all_human_responses,\n",
" use_aggregator=True,\n",
" use_stemmer=True,\n",
")\n",
"finetuned_bleu = bleu.compute(\n",
" predictions=all_finetuned_responses,\n",
" references=[[ref] for ref in all_human_responses]\n",
")\n",
"finetuned_fuzzy = compute_fuzzy_match(all_finetuned_responses, all_human_responses)\n",
"finetuned_exact = compute_exact_match(all_finetuned_responses, all_human_responses)\n",
"\n",
"print(\"\\n\" + \"=\" * 100)\n",
"print(\"Evaluation Metrics:\")\n",
"print(\"=\" * 100)\n",
"print(\"ORIGINAL MODEL:\")\n",
"print(f\" ROUGE: {orig_rouge}\")\n",
"print(f\" BLEU: {orig_bleu}\")\n",
"print(f\" Fuzzy Match Score: {orig_fuzzy:.2f}%\")\n",
"print(f\" Exact Match Accuracy: {orig_exact:.2f}%\\n\")\n",
"print(\"FINE-TUNED MODEL:\")\n",
"print(f\" ROUGE: {finetuned_rouge}\")\n",
"print(f\" BLEU: {finetuned_bleu}\")\n",
"print(f\" Fuzzy Match Score: {finetuned_fuzzy:.2f}%\")\n",
"print(f\" Exact Match Accuracy: {finetuned_exact:.2f}%\")\n",
"print(\"=\" * 100)\n",
"clear_memory()"
]
},
{
"cell_type": "code",
"execution_count": 15,
"id": "462546a7-6928-4723-b00e-23c3a4091d99",
"metadata": {},
"outputs": [
{
"name": "stderr",
"output_type": "stream",
"text": [
"2025-03-19 16:51:05,225 - INFO - Running inference with deterministic decoding and beam search.\n"
]
},
{
"name": "stdout",
"output_type": "stream",
"text": [
"Prompt:\n",
"Context:\n",
"CREATE TABLE customers (id INT PRIMARY KEY, name VARCHAR(100), country VARCHAR(50)); CREATE TABLE orders (order_id INT PRIMARY KEY, customer_id INT, total_amount DECIMAL(10,2), order_date DATE, FOREIGN KEY (customer_id) REFERENCES customers(id)); INSERT INTO customers (id, name, country) VALUES (1, 'Alice', 'USA'), (2, 'Bob', 'UK'), (3, 'Charlie', 'Canada'), (4, 'David', 'USA'); INSERT INTO orders (order_id, customer_id, total_amount, order_date) VALUES (101, 1, 500, '2024-01-15'), (102, 2, 300, '2024-01-20'), (103, 1, 700, '2024-02-10'), (104, 3, 450, '2024-02-15'), (105, 4, 900, '2024-03-05');\n",
"\n",
"Query:\n",
"Retrieve the total order amount for each customer, showing only customers from the USA, and sort the result by total order amount in descending order.\n",
"\n",
"Response:\n",
"SELECT customer_id, SUM(total_amount) as total_amount FROM orders JOIN customers ON orders.customer_id = customers.id WHERE customers.country = 'USA' GROUP BY customer_id ORDER BY total_amount DESC;\n"
]
}
],
"source": [
"import torch\n",
"from transformers import AutoModelForSeq2SeqLM, AutoTokenizer\n",
"import logging\n",
"\n",
"# Set up logging\n",
"logging.basicConfig(\n",
" level=logging.INFO,\n",
" format=\"%(asctime)s - %(levelname)s - %(message)s\",\n",
")\n",
"logger = logging.getLogger(__name__)\n",
"\n",
"# Ensure device is set (GPU if available)\n",
"device = torch.device(\"cuda\" if torch.cuda.is_available() else \"cpu\")\n",
"\n",
"# Load the fine-tuned model and tokenizer\n",
"model_name = \"text2sql_flant5base_finetuned\" \n",
"finetuned_model = AutoModelForSeq2SeqLM.from_pretrained(model_name, torch_dtype=torch.bfloat16)\n",
"tokenizer = AutoTokenizer.from_pretrained(\"google/flan-t5-base\")\n",
"finetuned_model.to(device)\n",
"\n",
"def run_inference(prompt_text: str) -> str:\n",
" \"\"\"\n",
" Runs inference on the fine-tuned model using deterministic decoding\n",
" with beam search, returning the generated SQL query.\n",
" \"\"\"\n",
" inputs = tokenizer(prompt_text, return_tensors=\"pt\").to(device)\n",
" generated_ids = finetuned_model.generate(\n",
" input_ids=inputs[\"input_ids\"],\n",
" max_new_tokens=100, # Adjust based on query complexity\n",
" temperature=0.1, # Deterministic output\n",
" num_beams=5, # Beam search for better output quality\n",
" early_stopping=True, # Stop early if possible\n",
" )\n",
" generated_sql = tokenizer.decode(generated_ids[0], skip_special_tokens=True)\n",
"\n",
" # Post-processing to remove repeated text\n",
" generated_sql = generated_sql.split(\";\")[0] + \";\" # Keep only the first valid SQL query\n",
"\n",
" return generated_sql\n",
"\n",
"# Sample context and query (example)\n",
"context = (\n",
" \"CREATE TABLE customers (id INT PRIMARY KEY, name VARCHAR(100), country VARCHAR(50)); \"\n",
" \"CREATE TABLE orders (order_id INT PRIMARY KEY, customer_id INT, total_amount DECIMAL(10,2), \"\n",
" \"order_date DATE, FOREIGN KEY (customer_id) REFERENCES customers(id)); \"\n",
" \"INSERT INTO customers (id, name, country) VALUES (1, 'Alice', 'USA'), (2, 'Bob', 'UK'), \"\n",
" \"(3, 'Charlie', 'Canada'), (4, 'David', 'USA'); \"\n",
" \"INSERT INTO orders (order_id, customer_id, total_amount, order_date) VALUES \"\n",
" \"(101, 1, 500, '2024-01-15'), (102, 2, 300, '2024-01-20'), \"\n",
" \"(103, 1, 700, '2024-02-10'), (104, 3, 450, '2024-02-15'), \"\n",
" \"(105, 4, 900, '2024-03-05');\"\n",
")\n",
"query = (\n",
" \"Retrieve the total order amount for each customer, showing only customers from the USA, \"\n",
" \"and sort the result by total order amount in descending order.\"\n",
")\n",
"\n",
"# Construct the prompt\n",
"sample_prompt = f\"\"\"Context:\n",
"{context}\n",
"\n",
"Query:\n",
"{query}\n",
"\n",
"Response:\n",
"\"\"\"\n",
"\n",
"logger.info(\"Running inference with deterministic decoding and beam search.\")\n",
"generated_sql = run_inference(sample_prompt)\n",
"\n",
"# Print output in the given format\n",
"print(\"Prompt:\")\n",
"print(\"Context:\")\n",
"print(context)\n",
"print(\"\\nQuery:\")\n",
"print(query)\n",
"print(\"\\nResponse:\")\n",
"print(generated_sql)\n"
]
},
{
"cell_type": "code",
"execution_count": 16,
"id": "a69f268e-bc69-4633-9c15-4e118c20178e",
"metadata": {},
"outputs": [
{
"name": "stdout",
"output_type": "stream",
"text": [
"✅ LoRA adapter saved at: text2sql_flant5base_finetuned\n",
"✅ Fully merged fine-tuned model saved at: text2sql_flant5base_finetuned_full\n"
]
}
],
"source": [
"import torch\n",
"import json\n",
"from transformers import AutoModelForSeq2SeqLM, AutoTokenizer\n",
"from peft import PeftModel\n",
"\n",
"# Define paths\n",
"base_model_name = \"google/flan-t5-base\" # Base model name\n",
"lora_model_path = \"text2sql_flant5base_finetuned\" # Folder where LoRA adapter is saved\n",
"full_model_output_path = \"text2sql_flant5base_finetuned_full\" # For merged full model\n",
"\n",
"# Load base model and tokenizer\n",
"base_model = AutoModelForSeq2SeqLM.from_pretrained(base_model_name, torch_dtype=torch.bfloat16)\n",
"tokenizer = AutoTokenizer.from_pretrained(base_model_name)\n",
"\n",
"# Load fine-tuned LoRA adapter model\n",
"lora_model = PeftModel.from_pretrained(base_model, lora_model_path)\n",
"\n",
"# ✅ Save the LoRA adapter separately (for users who want lightweight adapters)\n",
"lora_model.save_pretrained(lora_model_path)\n",
"tokenizer.save_pretrained(lora_model_path)\n",
"\n",
"# ✅ Merge LoRA into the base model to create a fully fine-tuned model\n",
"merged_model = lora_model.merge_and_unload()\n",
"\n",
"# ✅ Save the full fine-tuned model\n",
"merged_model.save_pretrained(full_model_output_path)\n",
"tokenizer.save_pretrained(full_model_output_path)\n",
"\n",
"# ✅ Save generation config (optional but recommended for inference settings)\n",
"generation_config = {\n",
" \"max_new_tokens\": 100,\n",
" \"temperature\": 0.1,\n",
" \"num_beams\": 5,\n",
" \"early_stopping\": True\n",
"}\n",
"with open(f\"{full_model_output_path}/generation_config.json\", \"w\") as f:\n",
" json.dump(generation_config, f)\n",
"\n",
"print(f\"✅ LoRA adapter saved at: {lora_model_path}\")\n",
"print(f\"✅ Fully merged fine-tuned model saved at: {full_model_output_path}\")\n"
]
},
{
"cell_type": "code",
"execution_count": 17,
"id": "f1c95dfc-6662-44d8-8ecc-bff414fecee5",
"metadata": {},
"outputs": [
{
"name": "stderr",
"output_type": "stream",
"text": [
"/venv/main/lib/python3.10/site-packages/transformers/generation/configuration_utils.py:629: UserWarning: `do_sample` is set to `False`. However, `temperature` is set to `0.1` -- this flag is only used in sample-based generation modes. You should set `do_sample=True` or unset `temperature`. This was detected when initializing the generation config instance, which means the corresponding file may hold incorrect parameterization and should be fixed.\n",
" warnings.warn(\n",
"/venv/main/lib/python3.10/site-packages/transformers/generation/configuration_utils.py:629: UserWarning: `do_sample` is set to `False`. However, `temperature` is set to `0.1` -- this flag is only used in sample-based generation modes. You should set `do_sample=True` or unset `temperature`.\n",
" warnings.warn(\n",
"2025-03-19 16:51:49,933 - INFO - Running inference with beam search decoding.\n"
]
},
{
"name": "stdout",
"output_type": "stream",
"text": [
"Prompt:\n",
"Context:\n",
"CREATE TABLE employees (id INT PRIMARY KEY, name VARCHAR(100), department VARCHAR(50), salary INT); CREATE TABLE projects (project_id INT PRIMARY KEY, project_name VARCHAR(100), budget INT); CREATE TABLE employee_projects (employee_id INT, project_id INT, role VARCHAR(50), FOREIGN KEY (employee_id) REFERENCES employees(id), FOREIGN KEY (project_id) REFERENCES projects(project_id)); INSERT INTO employees (id, name, department, salary) VALUES (1, 'Alice', 'Engineering', 90000), (2, 'Bob', 'Marketing', 70000), (3, 'Charlie', 'Engineering', 95000), (4, 'David', 'HR', 60000), (5, 'Eve', 'Engineering', 110000); INSERT INTO projects (project_id, project_name, budget) VALUES (101, 'AI Research', 500000), (102, 'Marketing Campaign', 200000), (103, 'Cloud Migration', 300000); INSERT INTO employee_projects (employee_id, project_id, role) VALUES (1, 101, 'Lead Engineer'), (2, 102, 'Marketing Specialist'), (3, 101, 'Engineer'), (4, 103, 'HR Coordinator'), (5, 101, 'AI Scientist');\n",
"\n",
"Query:\n",
"Find the names of employees who are working on the 'AI Research' project along with their roles.\n",
"\n",
"Response:\n",
"SELECT employees.name, employee_projects.role FROM employees INNER JOIN employee_projects ON employees.id = employee_projects.employee_id INNER JOIN projects ON employee_projects.project_id = projects.project_id WHERE projects.project_name = 'AI Research';\n"
]
}
],
"source": [
"import torch\n",
"from transformers import AutoModelForSeq2SeqLM, AutoTokenizer\n",
"import logging\n",
"\n",
"# Set up logging\n",
"logging.basicConfig(level=logging.INFO, format=\"%(asctime)s - %(levelname)s - %(message)s\")\n",
"logger = logging.getLogger(__name__)\n",
"\n",
"# Ensure device is set (GPU if available)\n",
"device = torch.device(\"cuda\" if torch.cuda.is_available() else \"cpu\")\n",
"\n",
"# Load the fine-tuned model and tokenizer\n",
"model_name = \"aarohanverma/text2sql-flan-t5-base-qlora-finetuned\"\n",
"model = AutoModelForSeq2SeqLM.from_pretrained(model_name, torch_dtype=torch.bfloat16).to(device)\n",
"tokenizer = AutoTokenizer.from_pretrained(\"aarohanverma/text2sql-flan-t5-base-qlora-finetuned\")\n",
"\n",
"# Ensure decoder start token is set\n",
"if model.config.decoder_start_token_id is None:\n",
" model.config.decoder_start_token_id = tokenizer.pad_token_id\n",
"\n",
"def run_inference(prompt_text: str) -> str:\n",
" \"\"\"\n",
" Runs inference on the fine-tuned model using beam search with fixes for repetition.\n",
" \"\"\"\n",
" inputs = tokenizer(prompt_text, return_tensors=\"pt\", truncation=True, max_length=512).to(device)\n",
"\n",
" generated_ids = model.generate(\n",
" input_ids=inputs[\"input_ids\"],\n",
" decoder_start_token_id=model.config.decoder_start_token_id, \n",
" max_new_tokens=100, \n",
" temperature=0.1, \n",
" num_beams=5, \n",
" repetition_penalty=1.2, \n",
" early_stopping=True, \n",
" )\n",
"\n",
" generated_sql = tokenizer.decode(generated_ids[0], skip_special_tokens=True)\n",
"\n",
" # Post-processing to remove repeated text\n",
" generated_sql = generated_sql.split(\";\")[0] + \";\" # Keep only the first valid SQL query\n",
"\n",
" return generated_sql\n",
"\n",
"# Example usage:\n",
"context = (\n",
" \"CREATE TABLE employees (id INT PRIMARY KEY, name VARCHAR(100), department VARCHAR(50), salary INT); \"\n",
" \"CREATE TABLE projects (project_id INT PRIMARY KEY, project_name VARCHAR(100), budget INT); \"\n",
" \"CREATE TABLE employee_projects (employee_id INT, project_id INT, role VARCHAR(50), \"\n",
" \"FOREIGN KEY (employee_id) REFERENCES employees(id), FOREIGN KEY (project_id) REFERENCES projects(project_id)); \"\n",
" \"INSERT INTO employees (id, name, department, salary) VALUES \"\n",
" \"(1, 'Alice', 'Engineering', 90000), (2, 'Bob', 'Marketing', 70000), \"\n",
" \"(3, 'Charlie', 'Engineering', 95000), (4, 'David', 'HR', 60000), (5, 'Eve', 'Engineering', 110000); \"\n",
" \"INSERT INTO projects (project_id, project_name, budget) VALUES \"\n",
" \"(101, 'AI Research', 500000), (102, 'Marketing Campaign', 200000), (103, 'Cloud Migration', 300000); \"\n",
" \"INSERT INTO employee_projects (employee_id, project_id, role) VALUES \"\n",
" \"(1, 101, 'Lead Engineer'), (2, 102, 'Marketing Specialist'), (3, 101, 'Engineer'), \"\n",
" \"(4, 103, 'HR Coordinator'), (5, 101, 'AI Scientist');\"\n",
")\n",
"\n",
"query = (\"Find the names of employees who are working on the 'AI Research' project along with their roles.\")\n",
"\n",
"\n",
"\n",
"# Construct the prompt\n",
"sample_prompt = f\"\"\"Context:\n",
"{context}\n",
"\n",
"Query:\n",
"{query}\n",
"\n",
"Response:\n",
"\"\"\"\n",
"\n",
"logger.info(\"Running inference with beam search decoding.\")\n",
"generated_sql = run_inference(sample_prompt)\n",
"\n",
"print(\"Prompt:\")\n",
"print(\"Context:\")\n",
"print(context)\n",
"print(\"\\nQuery:\")\n",
"print(query)\n",
"print(\"\\nResponse:\")\n",
"print(generated_sql)"
]
},
{
"cell_type": "code",
"execution_count": null,
"id": "97425ac4-ad46-4f38-b22d-071e161da20a",
"metadata": {},
"outputs": [],
"source": []
}
],
"metadata": {
"kernelspec": {
"display_name": "Python3 (ipykernel)",
"language": "python",
"name": "python3"
},
"language_info": {
"codemirror_mode": {
"name": "ipython",
"version": 3
},
"file_extension": ".py",
"mimetype": "text/x-python",
"name": "python",
"nbconvert_exporter": "python",
"pygments_lexer": "ipython3",
"version": "3.10.12"
}
},
"nbformat": 4,
"nbformat_minor": 5
}
|