File size: 56,670 Bytes
9e64e71 | 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 | {
"cells": [
{
"cell_type": "markdown",
"id": "c9e1656a",
"metadata": {},
"source": [
"# Prompting Baseline vs GRPO Comparison\\n\\nThis notebook compares five SQL agent methods on the same evaluation set: zero-shot, 1-shot, 3-shot, GRPO no-think, and GRPO thinking."
]
},
{
"cell_type": "markdown",
"id": "5d215e06",
"metadata": {},
"source": [
"## 1) Setup\\nDetect Colab, install dependencies when needed, and ensure Spider databases are available."
]
},
{
"cell_type": "code",
"execution_count": null,
"id": "0ff66f07",
"metadata": {},
"outputs": [],
"source": [
"import os\n",
"import subprocess\n",
"import sys\n",
"from pathlib import Path\n",
"\n",
"IN_COLAB = \"google.colab\" in sys.modules\n",
"\n",
"if IN_COLAB:\n",
" from google.colab import userdata\n",
"\n",
" token = userdata.get(\"GITHUB_TOKEN\")\n",
"\n",
" BRANCH = \"main\" # @param {type:\"string\"}\n",
" repo_url = f\"https://{token}@github.com/hjerpe/sql-env.git\"\n",
"\n",
" # Clone or update repo\n",
" if Path(\"sql-env\").exists():\n",
" subprocess.check_call([\"git\", \"-C\", \"sql-env\", \"pull\", \"-q\"])\n",
" else:\n",
" subprocess.check_call([\"git\", \"clone\", \"-q\", \"-b\", BRANCH, repo_url])\n",
" os.chdir(\"sql-env\")\n",
"\n",
" print(\"Colab detected: installing dependencies...\")\n",
" subprocess.check_call(\n",
" [sys.executable, \"-m\", \"pip\", \"install\", \"-q\", \"--upgrade\", \"pip\"]\n",
" )\n",
" # Uninstall vllm first to avoid version conflict with transformers 5.x\n",
" subprocess.call([sys.executable, \"-m\", \"pip\", \"uninstall\", \"-y\", \"vllm\"])\n",
" subprocess.check_call(\n",
" [\n",
" sys.executable,\n",
" \"-m\",\n",
" \"pip\",\n",
" \"install\",\n",
" \"-q\",\n",
" \"--no-deps\",\n",
" \"--force-reinstall\",\n",
" \".\",\n",
" ]\n",
" )\n",
" subprocess.check_call(\n",
" [\n",
" sys.executable,\n",
" \"-m\",\n",
" \"pip\",\n",
" \"install\",\n",
" \"-q\",\n",
" \"openenv-core[core]>=0.2.1\",\n",
" \"torch>=2.2.0\",\n",
" \"pandas>=2.0.0\",\n",
" \"matplotlib>=3.7.0\",\n",
" \"huggingface_hub>=0.37\",\n",
" \"jmespath\",\n",
" \"git+https://github.com/huggingface/transformers.git@main\",\n",
" ]\n",
" )\n",
" subprocess.check_call([sys.executable, \"scripts/download_spider_databases.py\"])\n",
" # Generate SFT data for few-shot examples\n",
" subprocess.check_call([sys.executable, \"scripts/generate_sft_data.py\"])\n",
"\n",
"project_root = Path.cwd().resolve()\n",
"if (\n",
" not (project_root / \"pyproject.toml\").exists()\n",
" and (project_root / \"sql-env\").exists()\n",
"):\n",
" project_root = (project_root / \"sql-env\").resolve()\n",
" os.chdir(project_root)\n",
"\n",
"if str(project_root) not in sys.path:\n",
" sys.path.insert(0, str(project_root))\n",
"\n",
"print(f\"Project root: {project_root}\")\n",
"print(f\"Running in Colab: {IN_COLAB}\")"
]
},
{
"cell_type": "markdown",
"id": "83a8bfbb",
"metadata": {},
"source": [
"## 2) Configuration"
]
},
{
"cell_type": "code",
"execution_count": null,
"id": "65decdf7",
"metadata": {},
"outputs": [],
"source": [
"N_EVAL_EPISODES = 50 # @param {type:\"integer\"}\n",
"STEP_BUDGET = 15\n",
"SEED = 42\n",
"\n",
"QUESTIONS_PATH = \"data/questions/questions_eval.json\"\n",
"DB_DIR = \"data/databases\"\n",
"\n",
"# ── Pick your base model ───────────────────────────────────────────\n",
"# Qwen3-0.6B → fits T4 (16 GB) comfortably\n",
"# Qwen3-1.7B → fits L4 (24 GB) with gradient checkpointing\n",
"BASE_MODEL_NAME = \"Qwen/Qwen3-0.6B\" # @param [\"Qwen/Qwen3-0.6B\", \"Qwen/Qwen3-1.7B\"]\n",
"\n",
"# ── Pick your trained checkpoints ─────────────────────────────────\n",
"# Should match the base model size you trained with.\n",
"# Set to \"none\" to skip that condition.\n",
"# Multiple GRPO checkpoints can be compared (e.g. v1 vs v2).\n",
"GRPO_MODEL_REPO = \"hjerpe/sqlenv-qwen3-0.6b-grpo\" # @param [\"hjerpe/sqlenv-qwen3-0.6b-grpo\", \"hjerpe/sqlenv-qwen3-1.7b-grpo\", \"none\"]\n",
"GRPO_V2_MODEL_REPO = \"none\" # @param [\"hjerpe/sqlenv-qwen3-0.6b-grpo-v2\", \"none\"]\n",
"GRPO_THINKING_MODEL_REPO = \"none\" # @param [\"hjerpe/sqlenv-qwen3-0.6b-grpo-think\", \"hjerpe/sqlenv-qwen3-1.7b-grpo-think\", \"none\"]\n",
"\n",
"print(f\"Base model: {BASE_MODEL_NAME}\")\n",
"print(f\"GRPO checkpoint (v1): {GRPO_MODEL_REPO}\")\n",
"print(f\"GRPO checkpoint (v2): {GRPO_V2_MODEL_REPO}\")\n",
"print(f\"GRPO thinking checkpoint: {GRPO_THINKING_MODEL_REPO}\")\n",
"print(f\"Eval episodes: {N_EVAL_EPISODES}, Step budget: {STEP_BUDGET}\")"
]
},
{
"cell_type": "markdown",
"id": "28bf9de3",
"metadata": {},
"source": [
"## 3) Imports"
]
},
{
"cell_type": "code",
"execution_count": null,
"id": "118aafdc",
"metadata": {},
"outputs": [],
"source": [
"from __future__ import annotations\n",
"\n",
"import gc\n",
"import json\n",
"import random\n",
"import re\n",
"from pathlib import Path\n",
"\n",
"import matplotlib.pyplot as plt\n",
"import pandas as pd\n",
"import torch\n",
"\n",
"from sql_env import SQLAction, SQLObservation\n",
"from sql_env.evaluation.policies import EvaluationResult, evaluate\n",
"from sql_env.server.sql_environment import SQLEnvironment\n",
"from sql_env.server.mock_tokenizer import MockTokenizer\n",
"from sql_env.training.data_loading import load_model_and_tokenizer\n",
"from sql_env.training.trl_adapter import get_tool_definitions\n",
"from scripts.generate_sft_data import get_system_prompt"
]
},
{
"cell_type": "markdown",
"id": "b6ed47e5",
"metadata": {},
"source": [
"## 4) Environment and Eval Data\\nCreate an environment instance and load the fixed evaluation set used by all comparison methods."
]
},
{
"cell_type": "code",
"execution_count": null,
"id": "00579e5b",
"metadata": {},
"outputs": [],
"source": [
"from sql_env.training.data_loading import validate_no_data_leak\n",
"\n",
"questions_path = Path(QUESTIONS_PATH)\n",
"db_dir = Path(DB_DIR)\n",
"\n",
"if not questions_path.exists():\n",
" raise FileNotFoundError(f\"Questions file not found: {questions_path}\")\n",
"\n",
"if not db_dir.exists():\n",
" print(\"Database directory missing, downloading Spider databases...\")\n",
" subprocess.check_call([sys.executable, \"scripts/download_spider_databases.py\"])\n",
"\n",
"# Guard against train/eval data leakage\n",
"train_path = Path(\"data/questions/questions_train.json\")\n",
"if train_path.exists():\n",
" validate_no_data_leak(str(train_path), str(questions_path))\n",
" print(\"Data leak check: PASSED (0 question overlap)\")\n",
"\n",
"env = SQLEnvironment(\n",
" questions_path=str(questions_path),\n",
" db_dir=str(db_dir),\n",
" tokenizer=MockTokenizer(),\n",
" step_budget=STEP_BUDGET,\n",
")\n",
"\n",
"with questions_path.open(\"r\", encoding=\"utf-8\") as handle:\n",
" eval_questions = json.load(handle)\n",
"\n",
"print(f\"Loaded {len(eval_questions)} eval questions from {questions_path}\")\n",
"print(f\"Environment ready with step budget = {STEP_BUDGET}\")"
]
},
{
"cell_type": "markdown",
"id": "84d2e123",
"metadata": {},
"source": [
"## 6) LLMToolCallingPolicy\n",
"Drive model inference using tool-calling chat templates with episode-aware history and parse-error fallback."
]
},
{
"cell_type": "code",
"execution_count": null,
"id": "814a69dc",
"metadata": {},
"outputs": [],
"source": [
"class LLMToolCallingPolicy:\n",
" \"\"\"Policy that mirrors TRL environment_factory rollouts exactly.\n",
"\n",
" TRL's rollout (see trl/trainer/grpo_trainer.py _tool_call_loop):\n",
" 1. Generate until EOS / max_new_tokens (no stop at </tool_call>).\n",
" 2. Parse ALL <tool_call> blocks from the completion.\n",
" 3. Append ONE assistant message with structured tool_calls list.\n",
" 4. Execute each call in order, appending one role:\"tool\" message each.\n",
" 5. Regenerate.\n",
"\n",
" The model never saw intermediate tool results while emitting a multi-call\n",
" turn during training — it committed to all N calls up-front. We buffer\n",
" parsed calls and drain them across N select_action invocations so the\n",
" per-step evaluation loop (evaluate() in policies.py) sees exactly the\n",
" same history the training rollout produced.\n",
" \"\"\"\n",
"\n",
" def __init__(\n",
" self,\n",
" model,\n",
" tokenizer,\n",
" tool_definitions: list[dict],\n",
" system_prompt: str,\n",
" few_shot_messages: list[dict] | None = None,\n",
" enable_thinking: bool = False,\n",
" max_new_tokens: int = 512,\n",
" verbose: bool = False,\n",
" ) -> None:\n",
" self.model = model\n",
" self.tokenizer = tokenizer\n",
" self.tool_definitions = tool_definitions\n",
" self.system_prompt = system_prompt\n",
" self.few_shot_messages = list(few_shot_messages or [])\n",
" self.enable_thinking = enable_thinking\n",
" self.max_new_tokens = max_new_tokens\n",
" self.verbose = verbose\n",
" self._messages: list[dict] = []\n",
" self._last_question: str | None = None\n",
" # Pending actions parsed from a single model generation, drained\n",
" # one-per-select_action. `pending_count` is the total number of\n",
" # tool calls that the model emitted in the current turn (so we\n",
" # can distinguish \"this is the last one\" from \"more to come\").\n",
" self._pending_actions: list[dict] = []\n",
" self._pending_count: int = 0\n",
" # True on the very first select_action of an episode, so we know\n",
" # NOT to append a tool message for a nonexistent previous action.\n",
" self._expect_tool_result: bool = False\n",
" # Logging counters\n",
" self.stats = {\n",
" \"total_calls\": 0, # select_action invocations\n",
" \"generations\": 0, # model.generate calls\n",
" \"parse_ok\": 0, # individual tool calls parsed from generations\n",
" \"parse_fail\": 0, # generations that produced zero parseable calls\n",
" \"multi_call_turns\": 0, # generations that produced >1 tool call\n",
" \"budget_exhaust\": 0,\n",
" \"parse_retries\": 0,\n",
" }\n",
" self.parse_errors: list[str] = []\n",
" self._current_question: str | None = None\n",
" self.failed_answers: list[dict] = []\n",
" self._episode_count = 0\n",
" self.reset()\n",
"\n",
" def reset(self) -> None:\n",
" self._messages = [{\"role\": \"system\", \"content\": self.system_prompt}]\n",
" self._messages.extend(self.few_shot_messages)\n",
" self._last_question = None\n",
" self._pending_actions = []\n",
" self._pending_count = 0\n",
" self._expect_tool_result = False\n",
"\n",
" def _to_device(self, value, device):\n",
" return value.to(device) if hasattr(value, \"to\") else value\n",
"\n",
" def _render_and_tokenize(self):\n",
" try:\n",
" rendered = self.tokenizer.apply_chat_template(\n",
" self._messages,\n",
" tools=self.tool_definitions,\n",
" add_generation_prompt=True,\n",
" tokenize=False,\n",
" enable_thinking=self.enable_thinking,\n",
" )\n",
" except TypeError:\n",
" rendered = self.tokenizer.apply_chat_template(\n",
" self._messages,\n",
" tools=self.tool_definitions,\n",
" add_generation_prompt=True,\n",
" tokenize=False,\n",
" )\n",
" encoded = self.tokenizer(rendered, return_tensors=\"pt\")\n",
" return encoded[\"input_ids\"], encoded.get(\"attention_mask\")\n",
"\n",
" def _parse_all_tool_calls(self, text: str) -> list[dict]:\n",
" \"\"\"Extract every <tool_call> JSON block from a completion.\n",
"\n",
" Returns a list of {\"name\": str, \"arguments\": dict} dicts. Invalid\n",
" blocks are silently skipped (they'd be dropped by TRL's parser too).\n",
" \"\"\"\n",
" if not text:\n",
" return []\n",
" pattern = re.compile(r\"<tool_call>\\s*(\\{.*?\\})\\s*</tool_call>\", re.DOTALL)\n",
" out: list[dict] = []\n",
" for raw_json in pattern.findall(text):\n",
" try:\n",
" obj = json.loads(raw_json)\n",
" except json.JSONDecodeError:\n",
" continue\n",
" if not isinstance(obj, dict):\n",
" continue\n",
" name = obj.get(\"name\")\n",
" args = obj.get(\"arguments\")\n",
" if isinstance(name, str) and isinstance(args, dict):\n",
" out.append({\"name\": name, \"arguments\": args})\n",
" return out\n",
"\n",
" def select_action(self, observation: SQLObservation) -> SQLAction:\n",
" self.stats[\"total_calls\"] += 1\n",
"\n",
" if observation.budget_remaining <= 1:\n",
" self.stats[\"budget_exhaust\"] += 1\n",
" # Flush any pending work; the episode is about to end anyway.\n",
" self._pending_actions = []\n",
" return SQLAction(action_type=\"ANSWER\", argument=\"budget_exhausted\")\n",
"\n",
" # New episode — reset message history and post the first user turn.\n",
" if observation.question != self._last_question:\n",
" self.reset()\n",
" self._last_question = observation.question\n",
" self._current_question = observation.question\n",
" self._episode_count += 1\n",
"\n",
" tables = []\n",
" for line in (observation.schema_info or \"\").split(\"\\n\"):\n",
" stripped = line.strip().lstrip(\"- \").strip()\n",
" if stripped and stripped.lower() != \"available tables:\":\n",
" tables.append(stripped)\n",
" # Matches TRL: reset() return string is appended to the last\n",
" # user message. See SQLEnvTRL.reset().\n",
" table_hint = (\n",
" f\"Tables: {', '.join(tables)}. \"\n",
" \"Use describe, sample, query, and answer tools.\"\n",
" )\n",
" self._messages.append(\n",
" {\n",
" \"role\": \"user\",\n",
" \"content\": f\"{observation.question}{table_hint}\",\n",
" }\n",
" )\n",
" self._expect_tool_result = False\n",
" elif self._expect_tool_result:\n",
" # The previous select_action returned an action; env.step just\n",
" # executed it and passed us its observation. Append the result\n",
" # as a role:\"tool\" message, matching TRL's per-call append.\n",
" result_text = observation.result or observation.error or \"\"\n",
" self._messages.append({\"role\": \"tool\", \"content\": result_text})\n",
" self._expect_tool_result = False\n",
"\n",
" # If we still have tool calls buffered from a multi-call generation,\n",
" # return the next one WITHOUT regenerating. This preserves training\n",
" # semantics: the model committed to all calls before seeing any\n",
" # results, so regenerating mid-batch would be a protocol violation.\n",
" if self._pending_actions:\n",
" tool_call = self._pending_actions.pop(0)\n",
" action = _tool_call_to_action(tool_call)\n",
" self._expect_tool_result = True\n",
" if self.verbose:\n",
" remaining = len(self._pending_actions)\n",
" tag = f\"[OK/buf+{remaining}]\" if remaining else \"[OK/buf]\"\n",
" print(f\" {tag} {action.action_type}: {str(action.argument)[:80]}\")\n",
" return action\n",
"\n",
" # Otherwise generate a fresh turn.\n",
" input_ids, attention_mask = self._render_and_tokenize()\n",
"\n",
" model_device = getattr(self.model, \"device\", None)\n",
" if model_device is None:\n",
" model_device = next(self.model.parameters()).device\n",
" input_ids = self._to_device(input_ids, model_device)\n",
" if attention_mask is not None:\n",
" attention_mask = self._to_device(attention_mask, model_device)\n",
"\n",
" if self.verbose and self.stats[\"generations\"] < 3:\n",
" print(\n",
" f\" [ctx] {input_ids.shape[-1]} input tokens, max_new={self.max_new_tokens}\"\n",
" )\n",
"\n",
" generate_kwargs = {\n",
" \"input_ids\": input_ids,\n",
" \"max_new_tokens\": self.max_new_tokens,\n",
" }\n",
" if attention_mask is not None:\n",
" generate_kwargs[\"attention_mask\"] = attention_mask\n",
"\n",
" with torch.no_grad():\n",
" output_ids = self.model.generate(**generate_kwargs)\n",
" self.stats[\"generations\"] += 1\n",
"\n",
" generated_ids = output_ids[0, input_ids.shape[-1] :]\n",
" generated_text = self.tokenizer.decode(\n",
" generated_ids, skip_special_tokens=True\n",
" ).strip()\n",
" generated_text_full = self.tokenizer.decode(\n",
" generated_ids, skip_special_tokens=False\n",
" ).strip()\n",
"\n",
" # Parse ALL tool calls from this generation (matches TRL). Try the\n",
" # skip-special-tokens=False variant as a fallback because sometimes\n",
" # the tool_call text sits next to a special token boundary.\n",
" parsed: list[dict] = []\n",
" for text_variant in (generated_text, generated_text_full):\n",
" parsed = self._parse_all_tool_calls(text_variant)\n",
" if parsed:\n",
" break\n",
"\n",
" if not parsed:\n",
" # Parse failure — model didn't emit a valid tool call.\n",
" self.stats[\"parse_fail\"] += 1\n",
" if len(self.parse_errors) < 5:\n",
" self.parse_errors.append(\n",
" f\"--- Parse error #{len(self.parse_errors) + 1} ---\\n\"\n",
" f\"text={generated_text[:300]}\"\n",
" )\n",
" if self.verbose:\n",
" print(f\" [PARSE FAIL] raw: {generated_text[:200]}\")\n",
"\n",
" # Don't end the episode on parse failure — append the\n",
" # failed output as an assistant message and let the\n",
" # episode continue. The evaluate() loop will call\n",
" # select_action again with the next observation.\n",
" # No extra coaching prompt — keeps format identical\n",
" # to what trained models see.\n",
" self._messages.append({\"role\": \"assistant\", \"content\": generated_text})\n",
" self.failed_answers.append(\n",
" {\n",
" \"question\": self._current_question,\n",
" \"raw_text\": generated_text[:500],\n",
" \"episode\": self._episode_count,\n",
" }\n",
" )\n",
" self._expect_tool_result = False\n",
" # Return a no-op DESCRIBE on the first available table\n",
" # so the episode doesn't end. The model wasted a turn\n",
" # but gets to keep exploring.\n",
" tables = []\n",
" for line in (observation.schema_info or \"\").split(\"\\n\"):\n",
" stripped = line.strip().lstrip(\"- \").strip()\n",
" if (\n",
" stripped\n",
" and stripped.lower() != \"available tables:\"\n",
" and \":\" not in stripped\n",
" ):\n",
" tables.append(stripped)\n",
" if tables:\n",
" return SQLAction(action_type=\"DESCRIBE\", argument=tables[0])\n",
" return SQLAction(action_type=\"ANSWER\", argument=\"parse_error\")\n",
"\n",
" # Append ONE assistant message containing the full tool_calls list\n",
" # (matches what TRL appends after a multi-call generation).\n",
" self._messages.append(\n",
" {\n",
" \"role\": \"assistant\",\n",
" \"tool_calls\": [\n",
" {\n",
" \"type\": \"function\",\n",
" \"function\": {\n",
" \"name\": tc[\"name\"],\n",
" \"arguments\": json.dumps(tc[\"arguments\"]),\n",
" },\n",
" }\n",
" for tc in parsed\n",
" ],\n",
" }\n",
" )\n",
" self.stats[\"parse_ok\"] += len(parsed)\n",
" if len(parsed) > 1:\n",
" self.stats[\"multi_call_turns\"] += 1\n",
"\n",
" # First call goes out now; the rest are buffered.\n",
" self._pending_actions = parsed[1:]\n",
" self._pending_count = len(parsed)\n",
" first = parsed[0]\n",
" action = _tool_call_to_action(first)\n",
" self._expect_tool_result = True\n",
" if self.verbose:\n",
" tag = f\"[OK/{len(parsed)}]\" if len(parsed) > 1 else \"[OK]\"\n",
" print(f\" {tag} {action.action_type}: {str(action.argument)[:80]}\")\n",
" return action\n",
"\n",
" def print_stats(self, label: str = \"\") -> None:\n",
" \"\"\"Print action statistics for debugging.\"\"\"\n",
" s = self.stats\n",
" gens = s[\"generations\"] or 1\n",
" print(f\"\\n{'=' * 60}\")\n",
" print(f\"Policy stats{f' ({label})' if label else ''}:\")\n",
" print(f\" select_action calls: {s['total_calls']}\")\n",
" print(f\" model.generate calls: {s['generations']}\")\n",
" print(f\" Tool calls parsed: {s['parse_ok']}\")\n",
" print(\n",
" f\" Multi-call turns: {s['multi_call_turns']} \"\n",
" f\"({s['multi_call_turns'] / gens:.0%} of generations)\"\n",
" )\n",
" print(\n",
" f\" Parse failures: {s['parse_fail']} \"\n",
" f\"({s['parse_fail'] / gens:.0%} of generations)\"\n",
" )\n",
" print(f\" Budget exhaust: {s['budget_exhaust']}\")\n",
" print(f\" Parse retries: {s['parse_retries']}\")\n",
" print(f\" Failed answer attempts logged: {len(self.failed_answers)}\")\n",
" if self.parse_errors:\n",
" print(f\"\\nFirst {len(self.parse_errors)} parse failure samples:\")\n",
" for sample in self.parse_errors:\n",
" print(sample)\n",
" print(f\"{'=' * 60}\")\n",
"\n",
"\n",
"def _tool_call_to_action(tool_call: dict) -> SQLAction:\n",
" \"\"\"Convert a parsed {name, arguments} dict into an SQLAction.\"\"\"\n",
" name = str(tool_call[\"name\"]).strip().lower()\n",
" arguments = tool_call[\"arguments\"]\n",
" if not isinstance(arguments, dict):\n",
" raise ValueError(\"Tool call arguments must be a dictionary\")\n",
"\n",
" if name == \"describe\":\n",
" argument = arguments.get(\"table_name\", arguments.get(\"table\"))\n",
" action_type = \"DESCRIBE\"\n",
" elif name == \"sample\":\n",
" argument = arguments.get(\"table_name\", arguments.get(\"table\"))\n",
" action_type = \"SAMPLE\"\n",
" elif name == \"query\":\n",
" argument = arguments.get(\"sql\")\n",
" action_type = \"QUERY\"\n",
" elif name == \"answer\":\n",
" argument = arguments.get(\"value\", arguments.get(\"answer\"))\n",
" action_type = \"ANSWER\"\n",
" else:\n",
" raise ValueError(f\"Unsupported tool name: {name}\")\n",
"\n",
" if argument is None:\n",
" raise ValueError(f\"Missing required argument for tool: {name}\")\n",
" return SQLAction(action_type=action_type, argument=str(argument))"
]
},
{
"cell_type": "markdown",
"id": "e04e13a4",
"metadata": {},
"source": "## 7) Few-Shot Example Builder\nBuild few-shot examples that demonstrate the complete tool-calling loop:\nquestion → describe → result → query → result → answer.\n\nExamples use the same message format as the evaluation policy (user/assistant roles,\n`<tool_call>` tags) so the model learns the exact pattern it needs to follow."
},
{
"cell_type": "code",
"execution_count": null,
"id": "8bc81b8d",
"metadata": {},
"outputs": [],
"source": [
"def build_few_shot_messages(\n",
" sft_path: str,\n",
" n_examples: int,\n",
" seed: int = 42,\n",
") -> list[dict]:\n",
" \"\"\"Build few-shot messages from SFT trajectories.\n",
"\n",
" SFT trajectories already use the exact training format\n",
" (user → assistant-with-tool_calls → tool → ...), so we just pass the\n",
" messages through verbatim. This guarantees zero drift between what the\n",
" model saw during SFT warmup / GRPO and what it sees in evaluation.\n",
" \"\"\"\n",
" if n_examples <= 0:\n",
" return []\n",
"\n",
" sft_file = Path(sft_path)\n",
" if not sft_file.exists():\n",
" raise FileNotFoundError(f\"SFT trajectories not found: {sft_file}\")\n",
"\n",
" with sft_file.open(\"r\", encoding=\"utf-8\") as f:\n",
" trajectories = json.load(f)\n",
"\n",
" # Prefer trajectories explicitly marked correct if the flag is present.\n",
" has_correct = any(\"correct\" in t for t in trajectories if isinstance(t, dict))\n",
" if has_correct:\n",
" candidates = [\n",
" t\n",
" for t in trajectories\n",
" if isinstance(t, dict)\n",
" and t.get(\"correct\") is True\n",
" and isinstance(t.get(\"messages\"), list)\n",
" ]\n",
" else:\n",
" candidates = [\n",
" t\n",
" for t in trajectories\n",
" if isinstance(t, dict) and isinstance(t.get(\"messages\"), list)\n",
" ]\n",
"\n",
" if not candidates:\n",
" print(\"Warning: no valid SFT trajectories found for few-shot examples\")\n",
" return []\n",
"\n",
" chosen = random.Random(seed).sample(candidates, min(n_examples, len(candidates)))\n",
"\n",
" few_shot: list[dict] = []\n",
" for traj in chosen:\n",
" for msg in traj[\"messages\"]:\n",
" role = msg.get(\"role\")\n",
" # Skip the system prompt — the policy adds its own.\n",
" if role == \"system\":\n",
" continue\n",
" if role in (\"user\", \"assistant\", \"tool\"):\n",
" few_shot.append(dict(msg))\n",
"\n",
" n_msgs = len(few_shot)\n",
" n_asst = sum(1 for m in few_shot if m.get(\"role\") == \"assistant\")\n",
" print(\n",
" f\"Few-shot: {len(chosen)} trajectories -> {n_msgs} messages ({n_asst} assistant turns)\"\n",
" )\n",
" return few_shot"
]
},
{
"cell_type": "markdown",
"id": "9aaeab16",
"metadata": {},
"source": [
"## 8) Base Model 3-Condition Evaluation\n",
"Load the base Qwen model once, build zero-shot/1-shot/3-shot policies, and run a fair comparison on the same eval split."
]
},
{
"cell_type": "code",
"execution_count": null,
"id": "5070d45a",
"metadata": {},
"outputs": [],
"source": [
"# ── Diagnostic: verify template + generation before full eval ──\n",
"# This cell runs ONE episode step to confirm the pipeline works end-to-end.\n",
"\n",
"_diag_model, _diag_tokenizer = load_model_and_tokenizer(BASE_MODEL_NAME)\n",
"if torch.cuda.is_available():\n",
" _diag_model = _diag_model.to(torch.device(\"cuda\"))\n",
"\n",
"_diag_prompt = get_system_prompt(enable_thinking=False)\n",
"_diag_tools = get_tool_definitions()\n",
"\n",
"# 1) Test template rendering\n",
"_diag_messages = [\n",
" {\"role\": \"system\", \"content\": _diag_prompt},\n",
" {\n",
" \"role\": \"user\",\n",
" \"content\": \"How many students are there?\\n\\nTables: students. Use describe, sample, query, and answer tools.\",\n",
" },\n",
"]\n",
"try:\n",
" _rendered = _diag_tokenizer.apply_chat_template(\n",
" _diag_messages,\n",
" tools=_diag_tools,\n",
" tokenize=False,\n",
" add_generation_prompt=True,\n",
" enable_thinking=False,\n",
" )\n",
" print(f\"Template OK ({len(_rendered)} chars)\")\n",
" print(f\"First 500 chars:\\n{_rendered[:500]}\")\n",
" print(f\"...\\nLast 200 chars:\\n{_rendered[-200:]}\")\n",
"except Exception as e:\n",
" print(f\"Template FAILED: {e}\")\n",
" _rendered = None\n",
"\n",
"# 2) Test generation\n",
"if _rendered:\n",
" _inputs = _diag_tokenizer(_rendered, return_tensors=\"pt\")\n",
" _inputs = {k: v.to(_diag_model.device) for k, v in _inputs.items()}\n",
" print(f\"\\nInput tokens: {_inputs['input_ids'].shape[-1]}\")\n",
"\n",
" with torch.no_grad():\n",
" _out = _diag_model.generate(**_inputs, max_new_tokens=200)\n",
" _new_ids = _out[0][_inputs[\"input_ids\"].shape[-1] :]\n",
" _text = _diag_tokenizer.decode(_new_ids, skip_special_tokens=True)\n",
" _text_full = _diag_tokenizer.decode(_new_ids, skip_special_tokens=False)\n",
" print(f\"\\nGenerated (skip_special=True):\\n{_text[:500]}\")\n",
" print(f\"\\nGenerated (skip_special=False):\\n{_text_full[:500]}\")\n",
" print(f\"\\nHas <tool_call>: {'<tool_call>' in _text_full}\")\n",
"\n",
"# 3) Test one full episode\n",
"print(\"\\n--- One episode test ---\")\n",
"_obs = env.reset(seed=42)\n",
"print(f\"Question: {_obs.question[:80]}\")\n",
"print(f\"Schema: {_obs.schema_info[:100]}\")\n",
"print(f\"Budget: {_obs.budget_remaining}, Done: {_obs.done}\")\n",
"\n",
"_policy = LLMToolCallingPolicy(\n",
" model=_diag_model,\n",
" tokenizer=_diag_tokenizer,\n",
" tool_definitions=_diag_tools,\n",
" system_prompt=_diag_prompt,\n",
" verbose=True,\n",
")\n",
"try:\n",
" _action = _policy.select_action(_obs)\n",
" print(f\"Action: {_action.action_type} = {_action.argument[:100]}\")\n",
"except Exception as e:\n",
" print(f\"select_action FAILED: {type(e).__name__}: {e}\")\n",
"\n",
"_policy.print_stats(\"diagnostic\")\n",
"\n",
"# Cleanup\n",
"del _diag_model, _diag_tokenizer, _policy\n",
"gc.collect()\n",
"if torch.cuda.is_available():\n",
" torch.cuda.empty_cache()\n",
"print(\"\\nDiagnostic complete.\")"
]
},
{
"cell_type": "code",
"execution_count": null,
"id": "faacf08b",
"metadata": {},
"outputs": [],
"source": [
"SFT_TRAJECTORIES_PATH = \"data/sft/sft_trajectories.json\"\n",
"\n",
"tool_definitions = get_tool_definitions()\n",
"system_prompt_nothink = get_system_prompt(enable_thinking=False)\n",
"\n",
"few_shot_1 = build_few_shot_messages(SFT_TRAJECTORIES_PATH, n_examples=1, seed=SEED)\n",
"few_shot_3 = build_few_shot_messages(SFT_TRAJECTORIES_PATH, n_examples=3, seed=SEED)\n",
"\n",
"print(f\"Loaded few-shot messages: 1-shot={len(few_shot_1)}, 3-shot={len(few_shot_3)}\")\n",
"\n",
"base_model, base_tokenizer = load_model_and_tokenizer(BASE_MODEL_NAME)\n",
"if torch.cuda.is_available():\n",
" base_model = base_model.to(torch.device(\"cuda\"))\n",
"print(f\"Loaded base model: {BASE_MODEL_NAME}\")"
]
},
{
"cell_type": "code",
"execution_count": null,
"id": "337bff7b",
"metadata": {},
"outputs": [],
"source": [
"def _progress(name: str):\n",
" def _callback(done: int, total: int) -> None:\n",
" if done == 1 or done == total or done % max(1, total // 5) == 0:\n",
" print(f\"[{name}] {done}/{total} episodes\")\n",
"\n",
" return _callback\n",
"\n",
"\n",
"# verbose=True on zero-shot so we can see raw model output for first episode\n",
"base_conditions = [\n",
" {\n",
" \"name\": \"zero-shot\",\n",
" \"policy\": LLMToolCallingPolicy(\n",
" model=base_model,\n",
" tokenizer=base_tokenizer,\n",
" tool_definitions=tool_definitions,\n",
" system_prompt=system_prompt_nothink,\n",
" few_shot_messages=None,\n",
" enable_thinking=False,\n",
" verbose=True,\n",
" ),\n",
" },\n",
" {\n",
" \"name\": \"1-shot\",\n",
" \"policy\": LLMToolCallingPolicy(\n",
" model=base_model,\n",
" tokenizer=base_tokenizer,\n",
" tool_definitions=tool_definitions,\n",
" system_prompt=system_prompt_nothink,\n",
" few_shot_messages=few_shot_1,\n",
" enable_thinking=False,\n",
" ),\n",
" },\n",
" {\n",
" \"name\": \"3-shot\",\n",
" \"policy\": LLMToolCallingPolicy(\n",
" model=base_model,\n",
" tokenizer=base_tokenizer,\n",
" tool_definitions=tool_definitions,\n",
" system_prompt=system_prompt_nothink,\n",
" few_shot_messages=few_shot_3,\n",
" enable_thinking=False,\n",
" ),\n",
" },\n",
"]"
]
},
{
"cell_type": "code",
"execution_count": null,
"id": "f4f9c222",
"metadata": {},
"outputs": [],
"source": [
"base_results: dict[str, EvaluationResult] = {}\n",
"base_policies: dict[str, LLMToolCallingPolicy] = {}\n",
"\n",
"for condition in base_conditions:\n",
" name = condition[\"name\"]\n",
" policy = condition[\"policy\"]\n",
" print(f\"\\nRunning condition: {name}\")\n",
" result = evaluate(\n",
" env,\n",
" policy,\n",
" n_episodes=N_EVAL_EPISODES,\n",
" seed=SEED,\n",
" progress_callback=_progress(name),\n",
" )\n",
" base_results[name] = result\n",
" base_policies[name] = policy\n",
" print(\n",
" f\"[{name}] accuracy={result.success_rate:.3f} avg_reward={result.avg_reward:.3f} \"\n",
" f\"avg_steps={result.avg_steps:.2f} completed={result.n_completed}/{result.n_episodes}\"\n",
" )\n",
" policy.print_stats(label=name)\n",
"\n",
" error_eps = [ep for ep in result.episodes if ep.error]\n",
" if error_eps:\n",
" print(f\" Episodes with errors: {len(error_eps)}\")\n",
" for ep in error_eps[:3]:\n",
" print(f\" ep#{ep.episode_index}: {ep.error[:120]}\")\n",
"\n",
" # Turn off verbose after first condition\n",
" policy.verbose = False\n",
"\n",
"print(\"\\nBase-model comparison complete.\")\n",
"print(f\"Collected results for: {', '.join(base_results.keys())}\")"
]
},
{
"cell_type": "markdown",
"id": "7988b07b",
"metadata": {},
"source": [
"## 9) GRPO Checkpoint Evaluation\n",
"Load GRPO checkpoints with graceful fallback so unavailable models do not block comparison runs."
]
},
{
"cell_type": "code",
"execution_count": null,
"id": "5e1ea7b1",
"metadata": {},
"outputs": [],
"source": [
"SYSTEM_PROMPT_THINK = get_system_prompt(enable_thinking=True)\n",
"\n",
"for var_name in (\"base_conditions\", \"base_model\", \"base_tokenizer\"):\n",
" if var_name in globals():\n",
" del globals()[var_name]\n",
"if torch.cuda.is_available():\n",
" torch.cuda.empty_cache()\n",
"gc.collect()\n",
"\n",
"grpo_conditions = [\n",
" {\n",
" \"name\": \"grpo-v1\",\n",
" \"repo_id\": GRPO_MODEL_REPO,\n",
" \"enable_thinking\": False,\n",
" \"system_prompt\": system_prompt_nothink,\n",
" },\n",
" {\n",
" \"name\": \"grpo-v2\",\n",
" \"repo_id\": GRPO_V2_MODEL_REPO,\n",
" \"enable_thinking\": False,\n",
" \"system_prompt\": system_prompt_nothink,\n",
" },\n",
" {\n",
" \"name\": \"grpo-thinking\",\n",
" \"repo_id\": GRPO_THINKING_MODEL_REPO,\n",
" \"enable_thinking\": True,\n",
" \"system_prompt\": SYSTEM_PROMPT_THINK,\n",
" },\n",
"]\n",
"\n",
"grpo_results: dict[str, EvaluationResult] = {}\n",
"grpo_policies: dict[str, LLMToolCallingPolicy] = {}\n",
"\n",
"for cfg in grpo_conditions:\n",
" if cfg[\"repo_id\"] == \"none\":\n",
" print(f\"\\nSkipping {cfg['name']} (set to 'none')\")\n",
" continue\n",
"\n",
" model = None\n",
" tokenizer = None\n",
" policy = None\n",
" print(f\"\\nLoading checkpoint: {cfg['repo_id']}\")\n",
" try:\n",
" model, tokenizer = load_model_and_tokenizer(cfg[\"repo_id\"])\n",
" except RuntimeError as exc:\n",
" print(f\"Warning: Could not load {cfg['repo_id']}. Skipping condition. ({exc})\")\n",
" continue\n",
"\n",
" try:\n",
" if torch.cuda.is_available():\n",
" model = model.to(torch.device(\"cuda\"))\n",
"\n",
" policy = LLMToolCallingPolicy(\n",
" model=model,\n",
" tokenizer=tokenizer,\n",
" tool_definitions=tool_definitions,\n",
" system_prompt=cfg[\"system_prompt\"],\n",
" few_shot_messages=None,\n",
" enable_thinking=cfg[\"enable_thinking\"],\n",
" verbose=True,\n",
" )\n",
"\n",
" result = evaluate(\n",
" env,\n",
" policy,\n",
" n_episodes=N_EVAL_EPISODES,\n",
" seed=SEED,\n",
" progress_callback=_progress(cfg[\"name\"]),\n",
" )\n",
" grpo_results[cfg[\"name\"]] = result\n",
" grpo_policies[cfg[\"name\"]] = policy\n",
" print(\n",
" f\"[{cfg['name']}] accuracy={result.success_rate:.3f} avg_reward={result.avg_reward:.3f} \"\n",
" f\"avg_steps={result.avg_steps:.2f} completed={result.n_completed}/{result.n_episodes}\"\n",
" )\n",
" policy.print_stats(label=cfg[\"name\"])\n",
"\n",
" error_eps = [ep for ep in result.episodes if ep.error]\n",
" if error_eps:\n",
" print(f\" Episodes with errors: {len(error_eps)}\")\n",
" for ep in error_eps[:3]:\n",
" print(f\" ep#{ep.episode_index}: {ep.error[:120]}\")\n",
" finally:\n",
" # Don't delete policy — we need it for analysis\n",
" if model is not None:\n",
" del model\n",
" if tokenizer is not None:\n",
" del tokenizer\n",
" gc.collect()\n",
" if torch.cuda.is_available():\n",
" torch.cuda.empty_cache()\n",
"\n",
"all_results: dict[str, EvaluationResult] = {**base_results, **grpo_results}\n",
"print(\"\\nGRPO checkpoint evaluation complete.\")\n",
"checkpoint_names = \", \".join(grpo_results.keys()) if grpo_results else \"none\"\n",
"print(f\"Checkpoint results collected for: {checkpoint_names}\")\n",
"print(f\"Total methods available for comparison: {', '.join(all_results.keys())}\")"
]
},
{
"cell_type": "markdown",
"id": "0428cd13",
"metadata": {},
"source": [
"## 10) Comparison Results\n\nThis section compares all available methods on the same evaluation subset:\n- **zero-shot**: base model, tool calling, no examples\n- **1-shot**: base model with one successful trajectory example\n- **3-shot**: base model with three successful trajectory examples\n- **grpo-no-think**: GRPO checkpoint without thinking mode\n- **grpo-thinking**: GRPO checkpoint with thinking mode\n\nIf a GRPO checkpoint is unavailable, that row is omitted automatically."
]
},
{
"cell_type": "code",
"execution_count": null,
"id": "edacc7b6",
"metadata": {},
"outputs": [],
"source": [
"def results_to_dataframe(\n",
" results: dict[str, EvaluationResult],\n",
" policies: dict[str, LLMToolCallingPolicy] | None = None,\n",
") -> pd.DataFrame:\n",
" \"\"\"Convert evaluation results + policy stats to a comparison DataFrame.\"\"\"\n",
" if not results:\n",
" return pd.DataFrame()\n",
"\n",
" ordered_names = [\n",
" \"zero-shot\",\n",
" \"1-shot\",\n",
" \"3-shot\",\n",
" \"grpo-v1\",\n",
" \"grpo-v2\",\n",
" \"grpo-thinking\",\n",
" ]\n",
"\n",
" rows = []\n",
" for name in list(ordered_names) + [n for n in results if n not in ordered_names]:\n",
" if name not in results:\n",
" continue\n",
" item = results[name]\n",
"\n",
" row = {\n",
" \"Method\": name,\n",
" \"Accuracy (%)\": round(item.success_rate * 100.0, 1),\n",
" \"Avg Reward\": round(item.avg_reward, 3),\n",
" \"Avg Steps\": round(item.avg_steps, 1),\n",
" }\n",
"\n",
" # Add policy stats if available\n",
" policy = (policies or {}).get(name)\n",
" if policy:\n",
" total = policy.stats[\"total_calls\"] or 1\n",
" row[\"Parse Rate (%)\"] = round(policy.stats[\"parse_ok\"] / total * 100, 1)\n",
" row[\"Parse Fails\"] = policy.stats[\"parse_fail\"]\n",
" row[\"Budget Exhaust\"] = policy.stats[\"budget_exhaust\"]\n",
"\n",
" row[\"Completed\"] = f\"{item.n_completed}/{item.n_episodes}\"\n",
" rows.append(row)\n",
"\n",
" return pd.DataFrame(rows)\n",
"\n",
"\n",
"def plot_comparison(df: pd.DataFrame) -> None:\n",
" \"\"\"Display grouped bar chart comparing accuracy and parse rate.\"\"\"\n",
" if df.empty or \"Accuracy (%)\" not in df.columns:\n",
" print(\"No results to plot.\")\n",
" return\n",
"\n",
" has_parse = \"Parse Rate (%)\" in df.columns\n",
" plot_df = df.sort_values(\"Accuracy (%)\", ascending=True).reset_index(drop=True)\n",
" colors_acc = [\n",
" \"#2b6cb0\" if not m.startswith(\"grpo\") else \"#2f855a\" for m in plot_df[\"Method\"]\n",
" ]\n",
"\n",
" if has_parse:\n",
" fig, (ax1, ax2) = plt.subplots(1, 2, figsize=(14, 5))\n",
" else:\n",
" fig, ax1 = plt.subplots(figsize=(10, 5))\n",
"\n",
" # Accuracy chart\n",
" bars = ax1.barh(plot_df[\"Method\"], plot_df[\"Accuracy (%)\"], color=colors_acc)\n",
" ax1.set_xlim(0, 100)\n",
" ax1.set_xlabel(\"Accuracy (%)\")\n",
" ax1.set_title(\"Answer Accuracy\")\n",
" ax1.grid(axis=\"x\", alpha=0.25)\n",
" for bar, value in zip(bars, plot_df[\"Accuracy (%)\"]):\n",
" ax1.text(\n",
" min(value + 1, 95),\n",
" bar.get_y() + bar.get_height() / 2,\n",
" f\"{value:.1f}%\",\n",
" va=\"center\",\n",
" )\n",
"\n",
" # Parse rate chart\n",
" if has_parse:\n",
" colors_parse = [\n",
" \"#805ad5\" if not m.startswith(\"grpo\") else \"#38a169\"\n",
" for m in plot_df[\"Method\"]\n",
" ]\n",
" bars2 = ax2.barh(\n",
" plot_df[\"Method\"], plot_df[\"Parse Rate (%)\"], color=colors_parse\n",
" )\n",
" ax2.set_xlim(0, 100)\n",
" ax2.set_xlabel(\"Parse Rate (%)\")\n",
" ax2.set_title(\"Tool-Call Format Compliance\")\n",
" ax2.grid(axis=\"x\", alpha=0.25)\n",
" for bar, value in zip(bars2, plot_df[\"Parse Rate (%)\"]):\n",
" ax2.text(\n",
" min(value + 1, 95),\n",
" bar.get_y() + bar.get_height() / 2,\n",
" f\"{value:.1f}%\",\n",
" va=\"center\",\n",
" )\n",
"\n",
" plt.tight_layout()\n",
" plt.show()\n",
"\n",
"\n",
"# Merge all policies for the table\n",
"all_policies = dict(base_policies)\n",
"if \"grpo_policies\" in dir() and grpo_policies:\n",
" all_policies.update(grpo_policies)\n",
"\n",
"comparison_df = results_to_dataframe(all_results, all_policies)\n",
"if comparison_df.empty:\n",
" print(\"No comparison results available. Run evaluation cells first.\")\n",
"else:\n",
" display(comparison_df)\n",
" plot_comparison(comparison_df)"
]
},
{
"cell_type": "markdown",
"id": "e0c352d8",
"metadata": {},
"source": [
"### Result Interpretation (Template)\n",
"\n",
"- Compare **zero-shot -> 1-shot -> 3-shot** to measure pure prompt-engineering gains.\n",
"- Compare **best prompting method vs GRPO checkpoints** to quantify training value.\n",
"- If only prompting rows are present, verify checkpoint availability and rerun GRPO cells.\n",
"- Keep `N_EVAL_EPISODES` and `SEED` fixed when comparing future runs for fairness."
]
},
{
"cell_type": "markdown",
"id": "6f374466",
"metadata": {},
"source": "### Failed Answer Analysis\nWhen the model fails to produce `<tool_call>` tags, it often still outputs an answer in natural language or bare JSON.\nThis analysis checks: **was the reasoning correct even when the format was wrong?**\n\nThis separates two distinct failure modes:\n- **Format failure only** — model knew the answer but didn't use `<tool_call>` tags\n- **Both format and reasoning failure** — model didn't know the answer either"
},
{
"cell_type": "code",
"execution_count": null,
"id": "2e1bee19",
"metadata": {},
"outputs": [],
"source": [
"from sql_env.server.verifier import verify_answer\n",
"\n",
"# Load gold answers for post-hoc checking\n",
"with questions_path.open(\"r\", encoding=\"utf-8\") as f:\n",
" _gold_data = json.load(f)\n",
"_gold_lookup = {q[\"question_text\"]: q for q in _gold_data}\n",
"\n",
"\n",
"def _extract_candidate_answer(raw_text: str) -> str | None:\n",
" \"\"\"Try to extract an answer value from unstructured model output.\"\"\"\n",
" text = raw_text.strip()\n",
"\n",
" # Try bare JSON: {\"name\": \"answer\", \"arguments\": {\"value\": \"...\"}}\n",
" try:\n",
" obj = json.loads(text)\n",
" if isinstance(obj, dict):\n",
" if \"value\" in obj:\n",
" return str(obj[\"value\"])\n",
" args = obj.get(\"arguments\", {})\n",
" if isinstance(args, dict) and \"value\" in args:\n",
" return str(args[\"value\"])\n",
" except json.JSONDecodeError:\n",
" pass\n",
"\n",
" # Try to find a short numeric or simple answer in the text\n",
" # Skip long explanatory text — only extract if it looks like a direct value\n",
" lines = [ln.strip() for ln in text.split(\"\\n\") if ln.strip()]\n",
" if len(lines) == 1 and len(lines[0]) < 50:\n",
" return lines[0]\n",
"\n",
" return None\n",
"\n",
"\n",
"def analyze_failed_answers(policy: LLMToolCallingPolicy, label: str) -> dict:\n",
" \"\"\"Check failed answer attempts against gold answers.\"\"\"\n",
" results = {\n",
" \"total_failures\": len(policy.failed_answers),\n",
" \"extractable\": 0,\n",
" \"correct_answer_wrong_format\": 0,\n",
" \"wrong_answer_wrong_format\": 0,\n",
" \"not_extractable\": 0,\n",
" \"examples_correct\": [],\n",
" \"examples_wrong\": [],\n",
" }\n",
"\n",
" for entry in policy.failed_answers:\n",
" question = entry[\"question\"]\n",
" raw_text = entry[\"raw_text\"]\n",
" gold = _gold_lookup.get(question)\n",
"\n",
" if gold is None:\n",
" continue\n",
"\n",
" candidate = _extract_candidate_answer(raw_text)\n",
" if candidate is None:\n",
" results[\"not_extractable\"] += 1\n",
" continue\n",
"\n",
" results[\"extractable\"] += 1\n",
" gold_answer = gold[\"gold_answer\"]\n",
" answer_type = gold.get(\"answer_type\", \"string\")\n",
"\n",
" try:\n",
" is_correct = verify_answer(candidate, gold_answer, answer_type)\n",
" except Exception:\n",
" is_correct = False\n",
"\n",
" if is_correct:\n",
" results[\"correct_answer_wrong_format\"] += 1\n",
" if len(results[\"examples_correct\"]) < 3:\n",
" results[\"examples_correct\"].append(\n",
" {\n",
" \"question\": question[:80],\n",
" \"predicted\": candidate[:60],\n",
" \"gold\": str(gold_answer)[:60],\n",
" }\n",
" )\n",
" else:\n",
" results[\"wrong_answer_wrong_format\"] += 1\n",
" if len(results[\"examples_wrong\"]) < 3:\n",
" results[\"examples_wrong\"].append(\n",
" {\n",
" \"question\": question[:80],\n",
" \"predicted\": candidate[:60],\n",
" \"gold\": str(gold_answer)[:60],\n",
" }\n",
" )\n",
"\n",
" return results\n",
"\n",
"\n",
"# Analyze all conditions\n",
"all_policies = dict(base_policies)\n",
"# Add GRPO policies if they were saved\n",
"if \"grpo_policies\" in dir():\n",
" all_policies.update(grpo_policies)\n",
"\n",
"analysis_rows = []\n",
"for name, policy in all_policies.items():\n",
" result = all_results.get(name)\n",
" analysis = analyze_failed_answers(policy, name)\n",
"\n",
" formal_correct = int(result.success_rate * result.n_completed) if result else 0\n",
" total_episodes = result.n_completed if result else 0\n",
"\n",
" print(f\"\\n{'=' * 60}\")\n",
" print(f\"Failed answer analysis: {name}\")\n",
" print(f\" Episodes: {total_episodes}\")\n",
" print(f\" Formally correct (format + answer): {formal_correct}\")\n",
" print(f\" Parse failures total: {analysis['total_failures']}\")\n",
" print(f\" - Extractable answer: {analysis['extractable']}\")\n",
" print(\n",
" f\" - Correct answer, wrong format: {analysis['correct_answer_wrong_format']}\"\n",
" )\n",
" print(\n",
" f\" - Wrong answer, wrong format: {analysis['wrong_answer_wrong_format']}\"\n",
" )\n",
" print(f\" - Not extractable: {analysis['not_extractable']}\")\n",
"\n",
" if analysis[\"examples_correct\"]:\n",
" print(\"\\n Examples — correct answer but wrong format:\")\n",
" for ex in analysis[\"examples_correct\"]:\n",
" print(f\" Q: {ex['question']}\")\n",
" print(f\" Predicted: {ex['predicted']} | Gold: {ex['gold']}\")\n",
"\n",
" # Count episodes where model answered in correct format but got wrong answer.\n",
" # Use episodes_with_failures (capped at total) not raw failure count.\n",
" episodes_with_parse_fail = (\n",
" min(\n",
" len([fa for fa in policy.failed_answers if fa.get(\"episode\", 0) > 0]),\n",
" total_episodes,\n",
" )\n",
" if hasattr(policy, \"failed_answers\")\n",
" else analysis[\"total_failures\"]\n",
" )\n",
" # Deduplicate by episode number\n",
" _fail_episodes = (\n",
" set(fa.get(\"episode\", 0) for fa in policy.failed_answers)\n",
" if hasattr(policy, \"failed_answers\")\n",
" else set()\n",
" )\n",
" episodes_with_parse_fail = len(_fail_episodes)\n",
" format_ok_answer_wrong = total_episodes - formal_correct - episodes_with_parse_fail\n",
" print(f\" Format OK, answer wrong: {format_ok_answer_wrong}\")\n",
" analysis_rows.append(\n",
" {\n",
" \"Method\": name,\n",
" \"Format+Answer OK\": formal_correct,\n",
" \"Format OK, Answer wrong\": format_ok_answer_wrong,\n",
" \"Format fail, Answer OK\": analysis[\"correct_answer_wrong_format\"],\n",
" \"Format fail, Answer wrong\": analysis[\"wrong_answer_wrong_format\"],\n",
" \"Not extractable\": analysis[\"not_extractable\"],\n",
" \"Total episodes\": total_episodes,\n",
" }\n",
" )\n",
"\n",
"print(f\"\\n{'=' * 60}\")\n",
"analysis_df = pd.DataFrame(analysis_rows)\n",
"if not analysis_df.empty:\n",
" display(analysis_df)"
]
}
],
"metadata": {
"kernelspec": {
"display_name": "Python 3 (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": "3.12.3"
}
},
"nbformat": 4,
"nbformat_minor": 5
}
|