Spaces:
Sleeping
Sleeping
File size: 58,371 Bytes
e60af4b | 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 | # /// script
# dependencies = [
# "requests",
# "huggingface_hub",
# ]
# ///
"""
=============================================================
SRE INCIDENT RESPONSE β COMPREHENSIVE TRAJECTORY COLLECTOR
=============================================================
Generates fine-tuning trajectories from the SRE incident simulator covering:
β’ All 10 tasks (8 training + 2 held-out compound scenarios)
β’ All 4 pools:
A β Phase 1 only (incident response)
B β Phase 2 only (code investigation, oracle belief injected)
C β Joint P1βP2 (full two-phase pipeline)
D β Held-out joint (generalization test)
β’ Full 17-action action space across both phases
β’ Multiple models from 1.5B to 70B+ (round-robin rotation)
β’ ALL episodes retained β negative-reward trajectories are kept as
hard-negative examples for RL/GRPO training
Output files:
sre_raw_trajectories.json β full episode records with score breakdowns
sre_sft_dataset.jsonl β per-step SFT samples (both phases, all rewards)
sre_grpo_dataset.jsonl β (prompt, chosen, rejected) pairs for GRPO/DPO
Usage:
export HF_TOKEN=hf_...
python sre_finetune_collector.py
Optional env vars:
NUM_EPISODES total episodes to collect (default: 200)
BASE_URL simulator URL (default: HF Space URL)
MAX_STEPS max steps per episode (default: 35)
SLEEP_BETWEEN seconds between steps (default: 0.6)
"""
from __future__ import annotations
import json
import os
import random
import time
import traceback
from collections import defaultdict
from typing import Any, Dict, List, Optional, Tuple
import requests
def upload_checkpoint(api, repo_id):
for fname in ["sre_raw_trajectories.jsonl", "sre_sft_dataset.jsonl", "sre_grpo_dataset.jsonl"]:
if os.path.exists(fname):
try:
api.upload_file(
path_or_fileobj=fname,
path_in_repo=fname,
repo_id=repo_id,
repo_type="dataset",
)
print(f"β
Uploaded {fname}")
except Exception as e:
print(f"β Upload failed {fname}: {e}")
# ββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
# Configuration
# ββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
HF_TOKEN = os.environ.get("HF_TOKEN")
BASE_URL = os.environ.get("BASE_URL", "https://meta-hf-hackathon-updated-policy.hf.space")
HF_ROUTER_URL = "https://router.huggingface.co/v1/chat/completions"
NUM_EPISODES = int(os.environ.get("NUM_EPISODES", "100"))
MAX_STEPS = int(os.environ.get("MAX_STEPS", "35"))
SLEEP_BETWEEN = float(os.environ.get("SLEEP_BETWEEN", "0.6"))
# ββ Model βββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
MODELS: List[str] = [
"Qwen/Qwen2.5-7B-Instruct:fastest",
]
# ββ Task registry per pool ββββββββββββββββββββββββββββββββββββββββββββββββββββ
# Pool A: P1-only incident response (all 8 training tasks)
# Pool B: P2-only code investigation (oracle belief injected; 7 tasks with code)
# Pool C: Joint P1βP2 full pipeline (all 8 training tasks)
# Pool D: Held-out joint (2 compound scenarios β generalization evaluation)
POOL_TASKS: Dict[str, List[str]] = {
"A": [
"memory_leak", "cascading_failure", "distributed_deadlock",
"circuit_breaker_noop", "aliased_fault", "severity_inversion",
"confidence_inversion", "info_ordering",
],
"B": [
"memory_leak", "cascading_failure", "distributed_deadlock",
"aliased_fault", "severity_inversion", "confidence_inversion", "info_ordering",
],
"C": [
"memory_leak", "cascading_failure", "distributed_deadlock",
"circuit_breaker_noop", "aliased_fault", "severity_inversion",
"confidence_inversion", "info_ordering",
],
"D": [
"heldout_aliased_severity", "heldout_confidence_ordering",
],
}
# Episode budget distribution across pools (must sum to 1.0)
POOL_WEIGHTS: Dict[str, float] = {"A": 0.35, "B": 0.20, "C": 0.35, "D": 0.10}
# ββ Action space definitions ββββββββββββββββββββββββββββββββββββββββββββββββββ
P1_DIAGNOSTIC = ["view_alerts", "query_logs", "check_metrics", "check_dependencies",
"check_deploy_history", "run_health_check"]
P1_REMEDIATION = ["restart_service", "rollback_deploy", "scale_service"]
P1_TERMINAL = ["declare_root_cause", "transition_to_phase2"]
P1_ACTIONS = P1_DIAGNOSTIC + P1_REMEDIATION + P1_TERMINAL
P2_DIAGNOSTIC = ["list_dir", "read_file", "search_code", "get_git_log", "get_file_diff"]
P2_TERMINAL = ["propose_patch", "declare_no_change"]
P2_ACTIONS = P2_DIAGNOSTIC + P2_TERMINAL
ALL_SERVICES = ["api_gateway", "auth", "orders", "payment", "cache", "database", "queue"]
TARGETED_ACTIONS = {
"query_logs", "check_metrics", "check_dependencies", "check_deploy_history",
"run_health_check", "restart_service", "rollback_deploy", "scale_service",
}
# Service dependency graph (for smarter fallbacks)
DEPENDENCY_GRAPH: Dict[str, List[str]] = {
"api_gateway": ["auth", "orders", "cache"],
"auth": ["database"],
"orders": ["database", "payment", "auth"],
"payment": ["queue", "database"],
"cache": [],
"database": [],
"queue": [],
}
# ββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
# System Prompts
# ββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
SYSTEM_PROMPT_P1 = """You are an expert SRE handling a production incident in a microservices system.
## Service Topology (downstream β upstream)
api_gateway β auth, orders, cache
auth β database
orders β database, payment, auth
payment β queue, database
cache, database, queue β (no dependencies)
## Phase 1 Action Space
Output EXACTLY ONE valid JSON action per turn. No markdown, no explanation.
Diagnostic (read-only):
{"action_type": "view_alerts"}
{"action_type": "query_logs", "target_service": "<svc>", "parameters": {"level": "ERROR", "keyword": "<optional>", "limit": 20}}
{"action_type": "check_metrics", "target_service": "<svc>"}
{"action_type": "check_dependencies", "target_service": "<svc>"}
{"action_type": "check_deploy_history", "target_service": "<svc>"}
{"action_type": "run_health_check", "target_service": "<svc>"}
Remediation (mutates state):
{"action_type": "restart_service", "target_service": "<svc>"}
{"action_type": "rollback_deploy", "target_service": "<svc>"}
{"action_type": "scale_service", "target_service": "<svc>", "parameters": {"replicas": 5}}
Declare root cause (ALL tasks β always call this once you have a diagnosis):
{"action_type": "declare_root_cause", "parameters": {"root_cause": "<specific diagnosis β service, what failed, why>"}}
Then for joint-mode tasks, ALSO transition to code investigation:
{"action_type": "transition_to_phase2", "parameters": {"belief": {
"suspected_service": "<root_cause_svc>",
"suspected_fault_class": "memory_leak|config_change|deadlock|dep_upgrade|none",
"service_confidence": 0.85,
"fault_confidence": 0.80,
"evidence_gaps": ["<what_you_didnt_check>"],
"estimated_p2_cost": "low|medium|high",
"decision": "transition",
"reasoning": "<concise evidence summary>"
}}}
## Investigation Strategy
1. ALWAYS start with view_alerts to understand severity and scope
2. check_metrics on the highest-alert service first
3. query_logs (level=ERROR) on degraded/down services
4. check_dependencies on the affected service to find upstream causes
5. check_deploy_history before any rollback
6. Remediate the ROOT CAUSE service, not the symptom
7. After 6-8 diagnostic steps you MUST call declare_root_cause with your diagnosis.
For P1-only tasks this ends the episode. For joint-mode tasks, follow it immediately
with transition_to_phase2. Do NOT keep diagnosing indefinitely β commit to a conclusion.
CRITICAL: Output ONLY valid JSON. No markdown. No explanation. No code blocks."""
SYSTEM_PROMPT_P2 = """You are an expert SRE investigating a code-level fault in a sandboxed repository.
## Phase 2 Action Space
Output EXACTLY ONE valid JSON action per turn. No markdown, no explanation.
Code Exploration:
{"action_type": "list_dir", "parameters": {"path": "."}}
{"action_type": "read_file", "parameters": {"path": "relative/path/to/file.py"}}
{"action_type": "search_code", "parameters": {"query": "<search string>", "file_pattern": "*.py", "max_hits": 20}}
{"action_type": "get_git_log", "parameters": {"path": ".", "n_commits": 15}}
{"action_type": "get_file_diff", "parameters": {"commit_sha": "<sha>", "path": "relative/path/file.py"}}
Terminal:
{"action_type": "propose_patch", "parameters": {"diff": "<unified diff β minimal, correct, applies cleanly>"}}
{"action_type": "declare_no_change", "parameters": {"reason": "<why no code fix is needed β infrastructure issue, not code>"}}
## Investigation Strategy
1. list_dir "." to understand project structure
2. get_git_log to find recent commits β especially the bad_commit_sha from Phase 1 context
3. get_file_diff on the suspicious commit SHA to see what changed
4. read_file on affected files to understand the bug
5. search_code to find related patterns or the fault injection site
6. If you found a code bug: propose_patch with a minimal, syntactically valid unified diff.
The bad_commit_sha in your context tells you exactly what changed β read that diff and revert/fix it.
7. declare_no_change ONLY if Phase 1 confirmed a spurious alert / circuit-breaker false positive
with no deployment or code change involved. If there IS a bad commit in the git log, propose_patch.
CRITICAL: Output ONLY valid JSON. No markdown. No explanation. No code blocks."""
# ββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
# Observation Formatters
# ββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
def _fmt_service_statuses(statuses: Dict[str, str]) -> str:
symbols = {"healthy": "β", "degraded": "~", "down": "β"}
return " ".join(
f"{symbols.get(v,'?')}{svc}({v})"
for svc, v in sorted(statuses.items())
)
def _fmt_action_result(result: Any, max_chars: int = 3000) -> str:
if result is None:
return "(no result)"
text = json.dumps(result, indent=2) if not isinstance(result, str) else result
if len(text) > max_chars:
text = text[:max_chars] + f"\n... [truncated {len(text)-max_chars} chars]"
return text
def format_initial_p1_obs(obs: dict, info: dict) -> str:
"""Format the very first observation for Phase 1."""
task = info.get("task_name", "unknown")
pool = info.get("pool", "?")
mode = info.get("mode", "unknown")
phase = obs.get("current_phase", 1)
svc_line = _fmt_service_statuses(obs.get("service_statuses", {}))
valid = obs.get("valid_actions", P1_ACTIONS)
return (
f"INCIDENT RESPONSE | Pool {pool} | Mode: {mode} | Task: {task}\n"
f"{'β'*60}\n"
f"Summary: {obs.get('incident_summary', 'No summary available')}\n"
f"Severity: {obs.get('severity', '?')} | "
f"Time Budget: {obs.get('time_budget_minutes', '?')} min | "
f"Max Steps: {obs.get('max_steps', MAX_STEPS)}\n"
f"Phase: {phase}\n"
f"\nService Statuses:\n {svc_line}\n"
f"Active Alerts: {obs.get('active_alerts_count', 0)}\n"
f"\nValid Actions: {valid}\n"
f"\nWhat is your FIRST action?"
)
def format_step_result_p1(obs: dict, reward: float) -> str:
"""Format a step result during Phase 1."""
svc_line = _fmt_service_statuses(obs.get("service_statuses", {}))
result = _fmt_action_result(obs.get("action_result"))
lines = [
f"Action Result (success={obs.get('action_success', '?')}): "
f"{obs.get('action_message', '')}",
f"\n{result}",
f"\n{'β'*40}",
f"Services: {svc_line}",
f"Alerts: {obs.get('active_alerts_count', 0)} active",
f"Step: {obs.get('steps_taken','?')}/{obs.get('max_steps', MAX_STEPS)} "
f"| Time: {obs.get('time_elapsed_minutes','?')}/{obs.get('time_budget_minutes','?')} min",
f"Reward: {reward:+.3f} | Cumulative: {obs.get('cumulative_reward', 0):+.3f}",
]
if obs.get("bad_commit_sha"):
lines.append(f"Bad Commit SHA: {obs['bad_commit_sha']} (remember for Phase 2)")
valid = obs.get("valid_actions", P1_ACTIONS)
lines.append(f"\nValid Actions: {valid}")
lines.append("\nWhat is your next action?")
return "\n".join(lines)
def format_initial_p2_obs(obs: dict, info: dict, belief: Optional[dict]) -> str:
"""Format the first Phase 2 observation (after transition or Pool B auto-start)."""
task = info.get("task_name", "unknown")
pool = info.get("pool", "?")
belief_text = ""
if belief:
belief_text = (
f"\n[Phase 1 Belief]\n"
f" Suspected service: {belief.get('suspected_service', '?')}\n"
f" Suspected fault: {belief.get('suspected_fault_class', '?')}\n"
f" Service confidence: {belief.get('service_confidence', 0):.0%}\n"
f" Fault confidence: {belief.get('fault_confidence', 0):.0%}\n"
f" Reasoning: {belief.get('reasoning', '')}\n"
f" P2 cost estimate: {belief.get('estimated_p2_cost', '?')}\n"
)
sha_line = (
f"Bad Commit SHA: {obs.get('bad_commit_sha', '(check git log)')}\n"
if obs.get("bad_commit_sha") else ""
)
return (
f"CODE INVESTIGATION | Pool {pool} | Task: {task}\n"
f"{'β'*60}\n"
f"{sha_line}"
f"{belief_text}\n"
f"Step: {obs.get('steps_taken', 0)}/{obs.get('max_steps', MAX_STEPS)} "
f"| Cumulative Reward: {obs.get('cumulative_reward', 0):+.3f}\n"
f"\nValid Actions: {obs.get('valid_actions', P2_ACTIONS)}\n"
f"\nWhat is your first Phase 2 action?"
)
def format_step_result_p2(obs: dict, reward: float) -> str:
"""Format a step result during Phase 2."""
result = _fmt_action_result(obs.get("action_result"))
lines = [
f"Action Result (success={obs.get('action_success', '?')}): "
f"{obs.get('action_message', '')}",
f"\n{result}",
f"\n{'β'*40}",
f"Step: {obs.get('steps_taken','?')}/{obs.get('max_steps', MAX_STEPS)}",
f"Reward: {reward:+.3f} | Cumulative: {obs.get('cumulative_reward', 0):+.3f}",
f"\nValid Actions: {obs.get('valid_actions', P2_ACTIONS)}",
"\nWhat is your next action?",
]
return "\n".join(lines)
# ββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
# Message Builder
# ββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
def build_messages(
history: List[Dict],
initial_user_msg: str,
phase: int,
max_recent: int = 10,
) -> List[Dict]:
"""
Build the full OpenAI-format messages list.
history: [{"action_json": str, "result_text": str, "reward": float}, ...]
max_recent caps how many turns are included to avoid context-length 422s.
"""
system = SYSTEM_PROMPT_P1 if phase == 1 else SYSTEM_PROMPT_P2
messages: List[Dict] = [
{"role": "system", "content": system},
{"role": "user", "content": initial_user_msg},
]
for entry in history[-max_recent:]:
messages.append({"role": "assistant", "content": entry["action_json"]})
messages.append({"role": "user", "content": entry["result_text"]})
return messages
# ββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
# Model Caller
# ββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
def call_model(
messages: List[Dict],
model: str,
temperature: float = 0.5,
max_tokens: int = 512,
retries: int = 3,
) -> str:
if not HF_TOKEN:
raise ValueError("HF_TOKEN is not set.")
payload = {
"model": model,
"messages": messages,
"max_tokens": max_tokens,
"temperature": temperature,
}
last_exc: Exception = RuntimeError("No attempts made")
for attempt in range(retries):
try:
resp = requests.post(
HF_ROUTER_URL,
headers={
"Authorization": f"Bearer {HF_TOKEN}",
"Content-Type": "application/json",
},
json=payload,
timeout=90,
)
resp.raise_for_status()
return resp.json()["choices"][0]["message"]["content"].strip()
except requests.HTTPError as e:
code = e.response.status_code if e.response is not None else 0
if code in (400, 422):
raise # client-format errors β retrying won't help; let caller handle
last_exc = e
wait = 2 ** attempt
print(f" [model retry {attempt+1}/{retries}] {e} β waiting {wait}s")
time.sleep(wait)
except Exception as e:
last_exc = e
wait = 2 ** attempt
print(f" [model retry {attempt+1}/{retries}] {e} β waiting {wait}s")
time.sleep(wait)
raise last_exc
def _merge_system_into_user(messages: List[Dict]) -> List[Dict]:
"""Fold system prompt into the first user message for models without system role."""
if not messages or messages[0]["role"] != "system":
return messages
system_text = messages[0]["content"]
rest = messages[1:]
if not rest or rest[0]["role"] != "user":
return rest
merged_first = {"role": "user", "content": f"{system_text}\n\n{rest[0]['content']}"}
return [merged_first] + rest[1:]
# Models confirmed to reject the system role β merged format used from the start.
_MODELS_NEEDING_MERGE: set = set()
def call_model_adaptive(
history: List[Dict],
initial_msg: str,
phase: int,
model: str,
temperature: float = 0.5,
) -> str:
"""
Call model with two layers of fallback:
400 (system role) β merge system into first user message and cache the result
so every subsequent step skips the wasted probe.
400 (after merge) β content still too long; halve history window.
422 (ctx length) β halve history window.
"""
use_merge = model in _MODELS_NEEDING_MERGE
probed_merge = use_merge # True = already confirmed in a prior step, no re-probe needed
max_recent = 10
while True:
messages = build_messages(history, initial_msg, phase, max_recent=max_recent)
if use_merge:
messages = _merge_system_into_user(messages)
try:
result = call_model(messages, model=model, temperature=temperature)
if use_merge and not probed_merge:
# Merged succeeded for the first time β cache it
_MODELS_NEEDING_MERGE.add(model)
print(f" [merged format confirmed for {model.split('/')[1]}, cached]")
return result
except requests.HTTPError as e:
code = e.response.status_code if e.response is not None else 0
if code == 400 and not use_merge:
use_merge = True
probed_merge = False
print(f" [400: probing merged format for {model.split('/')[1]}]")
elif code in (400, 422) and max_recent > 1:
max_recent = max(1, max_recent // 2)
print(f" [ctx truncated to last {max_recent} turns]")
else:
raise
# ββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
# Action Parsers
# ββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
def _extract_json(raw: str) -> dict:
"""Extract the first JSON object from model output and normalise colon-format action types.
Some models output {"action_type": "check_metrics:api_gateway"} instead of
the correct {"action_type": "check_metrics", "target_service": "api_gateway"}.
Split and normalise so the environment never sees an invalid action_type.
"""
start = raw.find("{")
end = raw.rfind("}") + 1
if start == -1 or end == 0:
raise ValueError("No JSON object in model output")
action = json.loads(raw[start:end])
atype = action.get("action_type", "")
if ":" in atype:
parts = atype.split(":", 1)
action["action_type"] = parts[0]
if parts[1] in ALL_SERVICES and "target_service" not in action:
action["target_service"] = parts[1]
return action
def _recent_sigs(recent_actions: List[dict], n: int = 3) -> set:
return {(a.get("action_type"), a.get("target_service")) for a in recent_actions[-n:]}
def _diversify_p1(obs: dict, recent_actions: List[dict]) -> dict:
"""Return the next logical diagnostic action that hasn't been done recently."""
statuses = obs.get("service_statuses") or {}
bad_svcs = [s for s, st in statuses.items() if st != "healthy"]
used_sigs = _recent_sigs(recent_actions, n=4)
used_svcs = {a.get("target_service") for a in recent_actions[-6:]}
used_types = {a.get("action_type") for a in recent_actions[-4:]}
# Build a uniform list of (score, action) tuples
candidates: List[Tuple[int, dict]] = []
for atype in P1_DIAGNOSTIC:
if atype == "view_alerts":
score = 0 if ("view_alerts", None) in used_sigs else 2
candidates.append((score, {"action_type": "view_alerts"}))
continue
for svc in (bad_svcs or ALL_SERVICES):
a: dict = {"action_type": atype, "target_service": svc}
if atype == "query_logs":
a["parameters"] = {"level": "ERROR", "limit": 20}
already_used = (atype, svc) in used_sigs
score = (not already_used) * 2 + (svc not in used_svcs) + (atype not in used_types)
candidates.append((score, a))
candidates.sort(key=lambda x: -x[0])
if candidates:
return candidates[0][1]
# Last resort
svc = next((s for s in ALL_SERVICES if s not in used_svcs), random.choice(ALL_SERVICES))
return {"action_type": "query_logs", "target_service": svc,
"parameters": {"level": "ERROR", "limit": 20}}
def parse_p1_action(raw: str, step: int, obs: dict, recent_actions: Optional[List[dict]] = None) -> dict:
"""Parse Phase 1 action with smart fallbacks and anti-repetition."""
recent_actions = recent_actions or []
valid = set(obs.get("valid_actions") or P1_ACTIONS)
try:
action = _extract_json(raw)
atype = action.get("action_type", "")
if atype not in valid:
action = _diversify_p1(obs, recent_actions) if step > 0 else {"action_type": "view_alerts"}
atype = action["action_type"]
# Ensure target_service for targeted actions
if atype in TARGETED_ACTIONS:
if "target_service" not in action or action["target_service"] not in ALL_SERVICES:
svcs = obs.get("available_services") or ALL_SERVICES
action["target_service"] = random.choice(svcs)
# Anti-repetition: if this exact (type, service) was used recently, diversify
sig = (action.get("action_type"), action.get("target_service"))
if sig in _recent_sigs(recent_actions, n=2) and atype not in ("declare_root_cause", "transition_to_phase2"):
action = _diversify_p1(obs, recent_actions)
atype = action["action_type"]
# Validate transition_to_phase2 belief structure
if atype == "transition_to_phase2":
params = action.setdefault("parameters", {})
belief = params.setdefault("belief", {})
degraded = [s for s, st in (obs.get("service_statuses") or {}).items()
if st != "healthy"]
belief.setdefault("suspected_service", degraded[0] if degraded else random.choice(ALL_SERVICES))
belief.setdefault("suspected_fault_class", "memory_leak")
belief.setdefault("service_confidence", 0.7)
belief.setdefault("fault_confidence", 0.65)
belief.setdefault("evidence_gaps", [])
belief.setdefault("estimated_p2_cost", "medium")
belief.setdefault("decision", "transition")
belief.setdefault("reasoning", "Transitioning based on collected evidence")
return action
except Exception:
if step == 0:
return {"action_type": "view_alerts"}
return _diversify_p1(obs, recent_actions)
def _force_p1_terminal(obs: dict) -> dict:
"""Build a best-effort terminal action from observed state."""
valid = set(obs.get("valid_actions") or P1_ACTIONS)
statuses = obs.get("service_statuses") or {}
degraded = [s for s, st in statuses.items() if st != "healthy"]
if "transition_to_phase2" in valid:
svc = degraded[0] if degraded else random.choice(ALL_SERVICES)
return {
"action_type": "transition_to_phase2",
"parameters": {"belief": {
"suspected_service": svc,
"suspected_fault_class": "memory_leak",
"service_confidence": 0.5,
"fault_confidence": 0.5,
"evidence_gaps": ["forced_terminal_after_step_limit"],
"estimated_p2_cost": "medium",
"decision": "transition",
"reasoning": f"Forced transition: degraded={degraded}",
}},
}
cause = (f"Degradation detected in: {', '.join(degraded)}"
if degraded else "Root cause undetermined within step budget")
return {"action_type": "declare_root_cause",
"parameters": {"root_cause": cause}}
def parse_p2_action(raw: str, step: int, obs: dict, recent_actions: Optional[List[dict]] = None) -> dict:
"""Parse Phase 2 action with smart fallbacks and anti-repetition."""
recent_actions = recent_actions or []
valid = set(obs.get("valid_actions") or P2_ACTIONS)
used_sigs = _recent_sigs(recent_actions, n=3)
p2_fallback_sequence = [
{"action_type": "list_dir", "parameters": {"path": "."}},
{"action_type": "get_git_log", "parameters": {"path": ".", "n_commits": 15}},
{"action_type": "search_code", "parameters": {"query": "error", "file_pattern": "*.py", "max_hits": 15}},
{"action_type": "search_code", "parameters": {"query": "def ", "file_pattern": "*.py", "max_hits": 10}},
{"action_type": "list_dir", "parameters": {"path": "src"}},
]
try:
action = _extract_json(raw)
atype = action.get("action_type", "")
if atype not in valid:
action = p2_fallback_sequence[step % len(p2_fallback_sequence)]
atype = action["action_type"]
# Ensure required params
params = action.setdefault("parameters", {})
if atype == "list_dir":
params.setdefault("path", ".")
elif atype == "read_file":
params.setdefault("path", ".")
elif atype == "search_code":
params.setdefault("query", "error")
params.setdefault("file_pattern", "*.py")
params.setdefault("max_hits", 15)
elif atype == "get_git_log":
params.setdefault("path", ".")
params.setdefault("n_commits", 10)
elif atype == "get_file_diff":
sha = obs.get("bad_commit_sha") or "HEAD"
params.setdefault("commit_sha", sha)
params.setdefault("path", ".")
elif atype == "propose_patch" and "diff" not in params:
action = {"action_type": "declare_no_change",
"parameters": {"reason": "Unable to determine code fix from available evidence"}}
elif atype == "declare_no_change":
params.setdefault("reason", "No code-level fix required based on investigation")
# Anti-repetition for non-terminal actions
sig = (action.get("action_type"), str(action.get("parameters", {}).get("path", "")))
if sig in used_sigs and atype not in P2_TERMINAL:
action = p2_fallback_sequence[(step + len(recent_actions)) % len(p2_fallback_sequence)]
return action
except Exception:
return p2_fallback_sequence[step % len(p2_fallback_sequence)]
# ββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
# Environment HTTP Helpers
# ββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
def _mask_p1_obs(obs: dict, pool: str) -> dict:
"""Pool A is p1_only β remove transition_to_phase2 the server incorrectly exposes."""
if pool == "A" and obs.get("valid_actions"):
obs = dict(obs)
obs["valid_actions"] = [a for a in obs["valid_actions"] if a != "transition_to_phase2"]
return obs
def env_reset(task_name: str, pool: str, seed: Optional[int] = None) -> dict:
body: dict = {"task_name": task_name, "pool": pool}
if seed is not None:
body["seed"] = seed
resp = requests.post(f"{BASE_URL}/reset", json=body, timeout=30)
resp.raise_for_status()
return resp.json()
def env_step(action: dict) -> dict:
resp = requests.post(f"{BASE_URL}/step", json=action, timeout=30)
resp.raise_for_status()
return resp.json()
def env_score(declared_patch: Optional[str], declared_no_change: bool,
belief_history: List[dict]) -> dict:
"""Fetch unified grader scores for the completed episode."""
try:
resp = requests.post(
f"{BASE_URL}/score",
json={
"declared_patch": declared_patch,
"declared_no_change": declared_no_change,
"belief_history": belief_history,
},
timeout=30,
)
resp.raise_for_status()
return resp.json()
except Exception as e:
print(f" [score] {e}")
return {}
def env_get_trajectory() -> dict:
"""Fetch the full trajectory from the server."""
try:
resp = requests.get(f"{BASE_URL}/trajectory", timeout=30)
resp.raise_for_status()
return resp.json()
except Exception:
return {}
# ββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
# Episode Runner
# ββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
def run_episode(
task_name: str,
pool: str,
model: str,
episode_id: int,
seed: Optional[int] = None,
) -> dict:
"""
Run one full episode (Phase 1, Phase 2, or Joint) through the HTTP API.
Returns a rich episode record including:
- step records with (action, raw_model_output, observation, reward)
- final score breakdown from /score
- SFT-ready message sequences per step
"""
print(f"\n{'β'*60}")
print(f" Ep {episode_id+1:>3} | Pool {pool} | Task: {task_name}")
print(f" Model: {model}")
print(f"{'β'*60}")
reset_resp = env_reset(task_name, pool, seed)
obs = _mask_p1_obs(reset_resp.get("observation", {}), pool)
info = reset_resp.get("info", {})
initial_phase = obs.get("current_phase", 1)
# Tracks for the episode
p1_steps: List[dict] = []
p2_steps: List[dict] = []
belief_history: List[dict] = []
declared_patch: Optional[str] = None
declared_no_change: bool = False
# Conversation history per phase (for message building)
p1_history: List[dict] = []
p2_history: List[dict] = []
last_belief: Optional[dict] = None
# Recent actions (flattened, for anti-repetition)
recent_actions: List[dict] = []
consecutive_errors = [0] # consecutive model call failures
consecutive_negative = [0] # consecutive negative-reward steps (patience)
# Initial user messages
initial_p1_msg = format_initial_p1_obs(obs, info)
initial_p2_msg: Optional[str] = None # set on transition
current_phase = initial_phase
done = False
for step_idx in range(MAX_STEPS):
if done:
break
# Pool B: transition to phase 2 only if the env actually started in phase 1.
# If the env auto-transitioned during reset, current_phase is already 2 β skip.
if pool == "B" and current_phase == 1 and len(p1_steps) == 0 and "transition_to_phase2" in (obs.get("valid_actions") or []):
raw = "{}"
action = {
"action_type": "transition_to_phase2",
"parameters": {"belief": {
"suspected_service": None,
"suspected_fault_class": None,
"service_confidence": 0.0,
"fault_confidence": 0.0,
"evidence_gaps": [],
"estimated_p2_cost": "unknown",
"decision": "transition",
"reasoning": "Pool B: oracle belief injected by environment",
}},
}
else:
# Hard ceiling: force terminal if too close to max_steps
p1_hard_limit = MAX_STEPS - 8 if pool in ("C", "D") else MAX_STEPS - 3
if current_phase == 1 and step_idx >= p1_hard_limit:
action = _force_p1_terminal(obs)
raw = json.dumps(action)
print(f" [step limit: forcing terminal]")
else:
# Call model with adaptive history truncation on 422
cur_history = p1_history if current_phase == 1 else p2_history
cur_initial = initial_p1_msg if current_phase == 1 else (
initial_p2_msg or format_initial_p2_obs(obs, info, last_belief)
)
if current_phase == 2 and initial_p2_msg is None:
initial_p2_msg = cur_initial
model_ok = True
try:
raw = call_model_adaptive(cur_history, cur_initial, current_phase, model)
except Exception as e:
print(f" [model error] {e}")
raw = "{}"
model_ok = False
# If model keeps failing in P1, force terminal after 8 consecutive errors
if not model_ok:
consecutive_errors[0] += 1
else:
consecutive_errors[0] = 0
if current_phase == 1 and consecutive_errors[0] >= 8:
action = _force_p1_terminal(obs)
raw = json.dumps(action)
consecutive_errors[0] = 0
print(f" [8 consecutive model errors: forcing terminal]")
elif current_phase == 1:
action = parse_p1_action(raw, step_idx, obs, recent_actions)
else:
action = parse_p2_action(raw, step_idx, obs, recent_actions)
print(f" step {step_idx+1:>2} | ph{current_phase} | {action.get('action_type')}"
+ (f"({action.get('target_service','')})" if action.get("target_service") else ""))
# Track terminal/transition actions before stepping
atype = action.get("action_type", "")
if atype == "transition_to_phase2":
belief = action.get("parameters", {}).get("belief", {})
last_belief = belief
belief_history.append(belief)
if atype == "propose_patch":
declared_patch = action.get("parameters", {}).get("diff", "")
if atype == "declare_no_change":
declared_no_change = True
# Step environment
try:
step_resp = env_step(action)
except Exception as e:
print(f" [env error] {e}")
break
reward = float(step_resp.get("reward", 0.0))
done = step_resp.get("done", False)
new_obs = step_resp.get("observation", {})
new_phase = new_obs.get("current_phase", current_phase)
print(f" reward={reward:+.3f} cumulative={new_obs.get('cumulative_reward', 0):+.3f}"
+ (" DONE" if done else ""))
# Build result text for next turn
if current_phase == 1:
result_text = format_step_result_p1(new_obs, reward)
else:
result_text = format_step_result_p2(new_obs, reward)
step_record = {
"step": step_idx,
"phase": current_phase,
"action": action,
"raw_output": raw,
"observation": new_obs,
"reward": reward,
"result_text": result_text, # stored for SFT building
}
if current_phase == 1:
p1_steps.append(step_record)
p1_history.append({"action_json": json.dumps(action), "result_text": result_text})
else:
p2_steps.append(step_record)
p2_history.append({"action_json": json.dumps(action), "result_text": result_text})
recent_actions.append(action)
# Patience: 10 consecutive negative rewards β force terminal immediately
if reward < 0:
consecutive_negative[0] += 1
else:
consecutive_negative[0] = 0
if consecutive_negative[0] >= 10 and not done:
print(f" [patience exhausted: 10 consecutive negatives β forcing terminal]")
if current_phase == 1:
term_action = _force_p1_terminal(new_obs)
else:
term_action = {"action_type": "declare_no_change",
"parameters": {"reason": "Patience exhausted β no progress detected"}}
try:
term_resp = env_step(term_action)
term_reward = float(term_resp.get("reward", 0.0))
done = term_resp.get("done", False)
term_obs = term_resp.get("observation", new_obs)
print(f" [forced terminal] reward={term_reward:+.3f} cumulative={term_obs.get('cumulative_reward',0):+.3f} DONE")
steps_list = p1_steps if current_phase == 1 else p2_steps
steps_list.append({"step": step_idx + 1, "phase": current_phase,
"action": term_action, "raw_output": "{}",
"observation": term_obs, "reward": term_reward,
"result_text": ""})
new_obs = term_obs
except Exception as e:
print(f" [forced terminal env error] {e}")
break
# Detect phase transition
if new_phase != current_phase and new_phase == 2:
print(" ββ Phase 1 β Phase 2 ββ")
initial_p2_msg = format_initial_p2_obs(new_obs, info, last_belief)
recent_actions.clear() # reset repetition tracking for the new phase
consecutive_negative[0] = 0 # reset patience on phase change
current_phase = new_phase
obs = _mask_p1_obs(new_obs, pool)
time.sleep(SLEEP_BETWEEN)
# Fetch unified scores
score = env_score(declared_patch, declared_no_change, belief_history)
cumulative = obs.get("cumulative_reward", 0.0)
print(f" Final cumulative reward: {cumulative:.3f}")
if score:
print(f" Scores: {json.dumps({k: round(v, 3) for k, v in score.items()})}")
return {
"episode_id": episode_id,
"task_name": task_name,
"pool": pool,
"model": model,
"seed": seed,
"p1_steps": p1_steps,
"p2_steps": p2_steps,
"num_p1_steps": len(p1_steps),
"num_p2_steps": len(p2_steps),
"cumulative_reward": round(cumulative, 4),
"score_breakdown": score,
"declared_patch": declared_patch,
"declared_no_change": declared_no_change,
"belief_history": belief_history,
"done": done,
# Reconstructed conversation contexts for SFT building
"_initial_p1_msg": initial_p1_msg,
"_initial_p2_msg": initial_p2_msg,
"_p1_history": p1_history,
"_p2_history": p2_history,
}
# ββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
# SFT Dataset Formatter
# ββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
def episode_to_sft_samples(ep: dict) -> List[dict]:
"""
Convert one episode into per-step SFT samples for BOTH phases.
ALL steps are included regardless of reward β negative-reward steps
provide hard-negative signal critical for RL/preference training.
The `reward` field is preserved so the training code can filter or weight.
"""
samples: List[dict] = []
def _extract_samples(steps, phase, initial_msg, history_key):
history_so_far: List[dict] = []
for i, step_rec in enumerate(steps):
system = SYSTEM_PROMPT_P1 if phase == 1 else SYSTEM_PROMPT_P2
messages = build_messages(history_so_far, initial_msg, phase=phase)
messages.append({
"role": "assistant",
"content": json.dumps(step_rec["action"]),
})
samples.append({
"messages": messages,
"reward": step_rec["reward"],
"phase": phase,
"action_type": step_rec["action"].get("action_type"),
"task_name": ep["task_name"],
"pool": ep["pool"],
"model": ep["model"],
"episode_id": ep["episode_id"],
"step": i,
})
history_so_far.append({
"action_json": json.dumps(step_rec["action"]),
"result_text": step_rec.get("result_text", ""),
})
if ep.get("p1_steps") and ep.get("_initial_p1_msg"):
_extract_samples(ep["p1_steps"], 1, ep["_initial_p1_msg"], "_p1_history")
if ep.get("p2_steps") and ep.get("_initial_p2_msg"):
_extract_samples(ep["p2_steps"], 2, ep["_initial_p2_msg"], "_p2_history")
return samples
# ββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
# GRPO / DPO Dataset Formatter
# ββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
def episodes_to_grpo_pairs(episodes: List[dict]) -> List[dict]:
"""
Build (prompt, chosen, rejected) triplets for GRPO/DPO training.
Three pairing strategies:
1. Within-episode: best vs worst step (same prompt context)
2. Cross-episode: same task+pool, different models, different outcomes
3. Phase-specific: separate within-phase pairs for P2
Chosen = action with higher reward.
Rejected = action with lower reward.
Both are kept regardless of absolute reward sign.
"""
pairs: List[dict] = []
# ββ Strategy 1: within-episode best/worst per phase βββββββββββββββββββββββ
for ep in episodes:
for phase, steps, initial_msg in [
(1, ep.get("p1_steps", []), ep.get("_initial_p1_msg", "")),
(2, ep.get("p2_steps", []), ep.get("_initial_p2_msg", "")),
]:
if len(steps) < 2 or not initial_msg:
continue
best = max(steps, key=lambda s: s["reward"])
worst = min(steps, key=lambda s: s["reward"])
if best["reward"] == worst["reward"]:
continue
if best is worst:
continue
prompt_msgs = build_messages([], initial_msg, phase=phase)
pairs.append({
"prompt": prompt_msgs,
"chosen": json.dumps(best["action"]),
"rejected": json.dumps(worst["action"]),
"chosen_reward": best["reward"],
"rejected_reward": worst["reward"],
"margin": best["reward"] - worst["reward"],
"task_name": ep["task_name"],
"pool": ep["pool"],
"phase": phase,
"strategy": "within_episode",
"episode_id": ep["episode_id"],
})
# ββ Strategy 2: cross-episode, same task+pool βββββββββββββββββββββββββββββ
by_task_pool: Dict[str, List[dict]] = defaultdict(list)
for ep in episodes:
key = f"{ep['task_name']}_{ep['pool']}"
by_task_pool[key].append(ep)
for key, task_eps in by_task_pool.items():
if len(task_eps) < 2:
continue
# Sort by cumulative reward; pair best vs worst episode
sorted_eps = sorted(task_eps, key=lambda e: e["cumulative_reward"])
best_ep = sorted_eps[-1]
worst_ep = sorted_eps[0]
if best_ep["cumulative_reward"] == worst_ep["cumulative_reward"]:
continue
if best_ep["episode_id"] == worst_ep["episode_id"]:
continue
# Use the first non-view_alerts action as representative
def _first_substantive_action(ep_inner, phase):
steps = ep_inner.get(f"p{phase}_steps", [])
for s in steps:
if s["action"].get("action_type") != "view_alerts":
return s
return steps[0] if steps else None
for phase in [1, 2]:
best_step = _first_substantive_action(best_ep, phase)
worst_step = _first_substantive_action(worst_ep, phase)
initial_msg = best_ep.get(f"_initial_p{phase}_msg", "")
if not best_step or not worst_step or not initial_msg:
continue
prompt_msgs = build_messages([], initial_msg, phase=phase)
pairs.append({
"prompt": prompt_msgs,
"chosen": json.dumps(best_step["action"]),
"rejected": json.dumps(worst_step["action"]),
"chosen_reward": best_ep["cumulative_reward"],
"rejected_reward": worst_ep["cumulative_reward"],
"margin": best_ep["cumulative_reward"] - worst_ep["cumulative_reward"],
"task_name": best_ep["task_name"],
"pool": best_ep["pool"],
"phase": phase,
"strategy": "cross_episode",
"best_model": best_ep["model"],
"worst_model": worst_ep["model"],
})
return pairs
# ββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
# Episode Schedule Builder
# ββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
def build_episode_schedule(n: int) -> List[Tuple[str, str, str, int]]:
"""
Return list of (task_name, pool, model, seed) tuples.
Distribution:
- Pools weighted by POOL_WEIGHTS
- Tasks within each pool: round-robin
- Models: round-robin across all MODELS
- Seeds: random per episode (for reproducibility, logged in output)
"""
schedule: List[Tuple[str, str, str, int]] = []
pool_counts = {
pool: max(1, round(n * weight))
for pool, weight in POOL_WEIGHTS.items()
}
# Adjust to exactly n
total = sum(pool_counts.values())
diff = n - total
if diff > 0:
pool_counts["C"] += diff
elif diff < 0:
pool_counts["A"] += diff # reduce A if over
model_idx = 0
for pool, count in pool_counts.items():
tasks = POOL_TASKS[pool]
for i in range(count):
task = tasks[i % len(tasks)]
model = MODELS[model_idx % len(MODELS)]
seed = random.randint(0, 99999)
schedule.append((task, pool, model, seed))
model_idx += 1
random.shuffle(schedule)
return schedule
# ββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
# Main
# ββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
def _flush_episode(ep: dict, raw_f, sft_f) -> Tuple[int, int, int]:
"""Append one episode to the open raw and SFT files. Returns (pos, zer, neg) step counts."""
clean = {k: v for k, v in ep.items() if not k.startswith("_")}
raw_f.write(json.dumps(clean) + "\n")
raw_f.flush()
samples = episode_to_sft_samples(ep)
for s in samples:
sft_f.write(json.dumps(s) + "\n")
sft_f.flush()
pos = sum(1 for s in samples if s["reward"] > 0)
zer = sum(1 for s in samples if s["reward"] == 0)
neg = sum(1 for s in samples if s["reward"] < 0)
return pos, zer, neg
def _finalize(all_episodes: List[dict], stats: Dict[str, List[float]]) -> None:
"""Generate GRPO pairs and print final statistics from whatever was collected."""
print(f"\n{'β'*60}")
print(f"β
Collected {len(all_episodes)} episodes")
grpo_pairs = episodes_to_grpo_pairs(all_episodes)
grpo_path = "sre_grpo_dataset.jsonl"
with open(grpo_path, "w") as f:
for p in grpo_pairs:
f.write(json.dumps(p) + "\n")
within = sum(1 for p in grpo_pairs if p["strategy"] == "within_episode")
cross = sum(1 for p in grpo_pairs if p["strategy"] == "cross_episode")
print(f"πΎ GRPO dataset ({len(grpo_pairs)} pairs) β {grpo_path}")
print(f" Pairs: {within} within-episode + {cross} cross-episode")
if not all_episodes:
return
all_rewards = [ep["cumulative_reward"] for ep in all_episodes]
print(f"\nπ Reward statistics:")
print(f" Overall: avg={sum(all_rewards)/len(all_rewards):.3f} "
f"max={max(all_rewards):.3f} min={min(all_rewards):.3f}")
print(f"\n By pool:")
for pool in ["A", "B", "C", "D"]:
rs = stats.get(f"pool_{pool}", [])
if rs:
print(f" Pool {pool}: n={len(rs):>3} avg={sum(rs)/len(rs):.3f} "
f"max={max(rs):.3f} min={min(rs):.3f}")
print(f"\n By task:")
for task in sorted(set(ep["task_name"] for ep in all_episodes)):
rs = stats.get(f"task_{task}", [])
if rs:
print(f" {task:<35} n={len(rs):>2} avg={sum(rs)/len(rs):.3f}")
print(f"\n By model tier:")
model_short_names = set()
for ep in all_episodes:
model_short_names.add(ep["model"].split("/")[1].split(":")[0])
for mname in sorted(model_short_names):
rs = stats.get(f"model_{mname}", [])
if rs:
print(f" {mname:<40} n={len(rs):>2} avg={sum(rs)/len(rs):.3f}")
def main():
from huggingface_hub import HfApi
api = HfApi(token=HF_TOKEN)
api.create_repo(repo_id="srinjoyd/sre-data", repo_type="dataset", exist_ok=True)
if not HF_TOKEN:
print("β HF_TOKEN is not set.\n export HF_TOKEN=hf_...")
return
print(f"π SRE Trajectory Collector")
print(f" Episodes: {NUM_EPISODES}")
print(f" Models: {len(MODELS)} (rotating)")
print(f" Tasks: {len(set(t for ts in POOL_TASKS.values() for t in ts))} unique")
print(f" Pools: A / B / C / D")
print(f" Base URL: {BASE_URL}")
print(f" Keeping ALL episodes (negative reward = hard negatives for RL)")
print(f" Saving incrementally β Ctrl+C safe\n")
schedule = build_episode_schedule(NUM_EPISODES)
all_episodes: List[dict] = []
stats: Dict[str, List[float]] = defaultdict(list)
total_pos = total_zer = total_neg = 0
raw_path = "sre_raw_trajectories.jsonl"
sft_path = "sre_sft_dataset.jsonl"
print(f"πΎ Writing to: {raw_path} | {sft_path} (appending per episode)")
print(f" GRPO pairs written at end (or on Ctrl+C)\n")
with open(raw_path, "a") as raw_f, open(sft_path, "a") as sft_f:
try:
for ep_id, (task, pool, model, seed) in enumerate(schedule):
try:
ep = run_episode(task, pool, model, ep_id, seed=seed)
except Exception as e:
print(f" [!] Episode {ep_id+1} FAILED: {e}")
traceback.print_exc()
time.sleep(2)
continue
all_episodes.append(ep)
pos, zer, neg = _flush_episode(ep, raw_f, sft_f)
upload_checkpoint(api, "srinjoyd/sre-data")
total_pos += pos
total_zer += zer
total_neg += neg
r = ep["cumulative_reward"]
stats[f"pool_{pool}"].append(r)
stats[f"task_{task}"].append(r)
stats[f"model_{model.split('/')[1].split(':')[0]}"].append(r)
print(f" [saved] ep {ep_id+1}/{NUM_EPISODES} | "
f"SFT steps so far: +{total_pos}/0:{total_zer}/-{total_neg}")
time.sleep(1.0)
except KeyboardInterrupt:
print(f"\n\nβ οΈ Interrupted after {len(all_episodes)} episodes β saving what we have...")
finally:
upload_checkpoint(api, "srinjoyd/sre-data") # always runs, even on crash
_finalize(all_episodes, stats)
print(f"\nπΎ Raw trajectories ({len(all_episodes)} eps) β {raw_path}")
print(f"πΎ SFT dataset ({total_pos+total_zer+total_neg} steps) β {sft_path}")
print(f" Reward split: +{total_pos} / 0:{total_zer} / -{total_neg}")
_finalize(all_episodes, stats)
print(f"\nπ‘ To upload to HuggingFace Hub:")
print(f" from datasets import Dataset")
print(f" import json")
print(f" sft = [json.loads(l) for l in open('sre_sft_dataset.jsonl')]")
print(f" grpo = [json.loads(l) for l in open('sre_grpo_dataset.jsonl')]")
print(f" Dataset.from_list(sft).push_to_hub('your-username/sre-sft-data')")
print(f" Dataset.from_list(grpo).push_to_hub('your-username/sre-grpo-data')")
if __name__ == "__main__":
main() |