Spaces:
Running
Running
File size: 26,898 Bytes
c745a99 | 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 | """Generate SFT dataset for the AWS RL environment.
Produces command-only assistant targets in chat-messages JSONL format,
ready for trl.SFTTrainer + peft LoRA.
Composition (by source):
55% success_first_step canonical command at step 1
20% multi_step_continuation step N>1 with prior command history
15% failure_recovery step 2 after a plausible typo/error
5% verification read-only check after task completion
5% hint_usage step 1 = aws help --task-hint
Tier sampling weights (warmup/beginner/intermediate/advanced/expert):
0.50 / 0.30 / 0.15 / 0.05 / 0.00
(expert skipped in SFT; GRPO will handle those with env reward)
Usage:
python data/build_sft_dataset.py --train 1500 --val 150 --reserve 200 --seed 42
"""
from __future__ import annotations
import argparse
import ast
import json
import random
import re
import textwrap
from collections import Counter
from pathlib import Path
from typing import Any, Callable
import yaml
def _find_repo_root(start: Path) -> Path:
"""Walk up from `start` looking for the dir that contains server/services/tasks/.
Makes the script location-independent: works whether it lives at repo root,
in data/, scripts/, or anywhere else in the tree.
"""
for p in [start, *start.parents]:
if (p / "server" / "services" / "tasks").is_dir():
return p
return start
REPO_ROOT = _find_repo_root(Path(__file__).resolve().parent)
TASKS_DIR = REPO_ROOT / "server" / "services" / "tasks"
TESTS_DIR = REPO_ROOT / "tests_tasks"
DEFAULT_OUT_DIR = REPO_ROOT / "data" / "sft"
TIERS = ["warmup", "beginner", "intermediate", "advanced", "expert"]
TIER_WEIGHTS = {
"warmup": 0.50,
"beginner": 0.30,
"intermediate": 0.15,
"advanced": 0.05,
"expert": 0.00,
}
SOURCE_MIX = {
"success_first_step": 0.55,
"multi_step_continuation": 0.20,
"failure_recovery": 0.15,
"verification": 0.05,
"hint_usage": 0.05,
}
TESTS_FILES = {
"warmup": ("test_warmup_tasks.py", "WARMUP_COMMANDS"),
"beginner": ("test_beginner_tasks.py", "BEGINNER_COMMANDS"),
"intermediate": ("test_intermediate_tasks.py", "INTERMEDIATE_COMMANDS"),
"advanced": ("test_advanced_tasks.py", "ADVANCED_COMMANDS"),
"expert": ("test_expert_tasks.py", "EXPERT_COMMANDS"),
}
SYSTEM_PROMPT = textwrap.dedent(
"""
You are an AWS cloud engineer interacting with a real AWS environment via CLI.
Each turn you must send exactly ONE valid AWS CLI command (starting with 'aws').
You will be given a task to accomplish. Read the task description carefully.
Use the command output and error messages to guide your next action.
Rules:
- Only send AWS CLI commands (e.g. 'aws s3 ls', 'aws dynamodb create-table ...')
- One command per turn — no pipes, no shell syntax, no chaining
- Reply with ONLY the command, nothing else — no explanations, no quotes
- If unsure, use 'aws help' to get unstuck, but try to be specific to the service if possible (e.g. 'aws s3 help')
- When ever you need a hint, use 'aws help --task-hint' to get a task-specific hint (you can use this multiple times for more hints, but hints reduce your reward)
"""
).strip()
# Plausible reset-state outputs the env might return. Weighted toward "" since
# that's the most likely real value; others add mild prompt variance so the
# SFT trainer sees diverse surface forms of the same logical state.
_INITIAL_OUTPUTS: list[tuple[str, float]] = [
("", 0.7),
("Environment reset. Infra state wiped.", 0.2),
("Environment ready.", 0.1),
]
def _sample_initial_output(rng: random.Random) -> str:
values, weights = zip(*_INITIAL_OUTPUTS)
return rng.choices(values, weights=weights, k=1)[0]
def load_tasks() -> dict[str, list[dict]]:
out: dict[str, list[dict]] = {}
for tier in TIERS:
path = TASKS_DIR / f"{tier}.yaml"
if not path.exists():
out[tier] = []
continue
with open(path) as f:
raw = yaml.safe_load(f) or []
tasks = [t for t in raw if isinstance(t, dict) and "task_id" in t and "description" in t]
out[tier] = tasks
return out
def load_canonical_commands() -> dict[int, list[str]]:
"""Parse command dicts from tests_tasks/*.py without executing the files.
Tests import pytest and set up fixtures, so importlib.exec fails outside
the venv. AST-based literal extraction avoids that: we find the
module-level `X_COMMANDS = {...}` assignment and literal-eval its value.
Entries with non-literal values (f-strings etc.) are skipped silently.
"""
out: dict[int, list[str]] = {}
for _tier, (fname, var) in TESTS_FILES.items():
path = TESTS_DIR / fname
if not path.exists():
continue
try:
tree = ast.parse(path.read_text())
except SyntaxError:
continue
for node in ast.iter_child_nodes(tree):
if isinstance(node, ast.AnnAssign) and isinstance(node.target, ast.Name):
target_id, value = node.target.id, node.value
elif isinstance(node, ast.Assign) and len(node.targets) == 1 and isinstance(node.targets[0], ast.Name):
target_id, value = node.targets[0].id, node.value
else:
continue
if target_id != var or value is None:
continue
try:
d = ast.literal_eval(value)
except (ValueError, SyntaxError):
continue
if not isinstance(d, dict):
continue
for tid, cmd in d.items():
if not isinstance(tid, int):
continue
if isinstance(cmd, str):
out[tid] = [cmd]
elif isinstance(cmd, (list, tuple)):
seq = [c for c in cmd if isinstance(c, str)]
if seq:
out[tid] = seq
return out
def task_has_dynamic_ids(cmd_seq: list[str]) -> bool:
"""Detect commands that reference runtime-resolved IDs (sg-, subnet-, etc.)."""
for cmd in cmd_seq:
if re.search(r"\b(sg|subnet|vpc|ami|rtb|eni|igw|nat|eip|snap|vol)-[a-f0-9]{8,}\b", cmd):
return True
if re.search(r"\bi-[a-f0-9]{8,}\b", cmd):
return True
return False
OP_OUTPUTS: dict[str, str] = {
"ls": "",
"list-buckets": '{"Buckets":[]}',
"list-tables": '{"TableNames":[]}',
"list-functions": '{"Functions":[]}',
"list-queues": "{}",
"list-topics": '{"Topics":[]}',
"list-users": '{"Users":[]}',
"list-secrets": '{"SecretList":[]}',
"list-clusters": '{"clusterArns":[]}',
"list-named-queries": '{"NamedQueryIds":[]}',
"describe-instances": '{"Reservations":[]}',
"describe-db-instances": '{"DBInstances":[]}',
"describe-cache-clusters": '{"CacheClusters":[]}',
"get-databases": '{"DatabaseList":[]}',
"create-bucket": '{"Location":"/<RESOURCE>"}',
"put-object": '{"ETag":"\\"d41d8cd98f00b204e9800998ecf8427e\\""}',
"put-bucket-versioning": "",
"put-bucket-policy": "",
"create-table": '{"TableDescription":{"TableName":"<RESOURCE>","TableStatus":"ACTIVE"}}',
"put-item": "{}",
"create-topic": '{"TopicArn":"arn:aws:sns:us-east-1:000000000000:<RESOURCE>"}',
"create-queue": '{"QueueUrl":"https://sqs.us-east-1.amazonaws.com/000000000000/<RESOURCE>"}',
"subscribe": '{"SubscriptionArn":"arn:aws:sns:us-east-1:000000000000:<RESOURCE>:abc123"}',
"create-role": '{"Role":{"RoleName":"<RESOURCE>","Arn":"arn:aws:iam::000000000000:role/<RESOURCE>"}}',
"attach-role-policy": "",
"create-function": '{"FunctionName":"<RESOURCE>","FunctionArn":"arn:aws:lambda:us-east-1:000000000000:function:<RESOURCE>"}',
"create-policy": '{"Policy":{"PolicyName":"<RESOURCE>","Arn":"arn:aws:iam::000000000000:policy/<RESOURCE>"}}',
"create-secret": '{"ARN":"arn:aws:secretsmanager:us-east-1:000000000000:secret:<RESOURCE>"}',
"get-bucket-policy": '{"Policy":"{\\"Version\\":\\"2012-10-17\\",\\"Statement\\":[{\\"Effect\\":\\"Allow\\",\\"Principal\\":\\"*\\",\\"Action\\":\\"s3:*\\",\\"Resource\\":\\"*\\"}]}"}',
}
_RESOURCE_FLAGS = (
"--bucket",
"--table-name",
"--role-name",
"--function-name",
"--queue-name",
"--topic-arn",
"--policy-name",
"--name",
"--secret-id",
"--cluster",
"--key",
)
def simulate_output(command: str) -> str:
tokens = command.split()
if len(tokens) < 3:
return ""
op = tokens[2]
resource = ""
for i, tok in enumerate(tokens):
if tok in _RESOURCE_FLAGS and i + 1 < len(tokens):
resource = tokens[i + 1]
break
template = OP_OUTPUTS.get(op, "")
return template.replace("<RESOURCE>", resource)
def _mistake_wrong_operation(cmd: str) -> tuple[str, str] | None:
swaps = [
("list-tables", "ls"),
("list-buckets", "ls-all"),
("describe-instances", "list-instances"),
("list-functions", "list-lambdas"),
("create-bucket", "make-bucket"),
("put-item", "insert-item"),
("create-table", "make-table"),
("get-bucket-policy", "show-bucket-policy"),
("attach-role-policy", "attach-policy"),
("create-role", "new-role"),
]
for good, bad in swaps:
if good in cmd:
return cmd.replace(good, bad, 1), (
f"aws: error: argument operation: Invalid choice: '{bad}'"
)
return None
def _mistake_missing_arg(cmd: str) -> tuple[str, str] | None:
m = re.search(r" (--[a-z-]+) (\S+)", cmd)
if not m:
return None
flag, value = m.group(1), m.group(2)
wrong = cmd.replace(f"{flag} {value}", "", 1).rstrip()
return wrong, f"aws: error: the following arguments are required: {flag}"
def _mistake_wrong_service(cmd: str) -> tuple[str, str] | None:
swaps = [
("aws dynamodb", "aws dynamo"),
("aws secretsmanager", "aws secrets"),
("aws cloudformation", "aws cfn"),
("aws elasticache", "aws elastic"),
("aws apigateway", "aws apigw"),
]
for good, bad in swaps:
if cmd.startswith(good):
wrong = cmd.replace(good, bad, 1)
return wrong, (
f"aws: error: argument command: Invalid choice: '{bad.split()[-1]}'"
)
return None
def _mistake_s3_vs_s3api(cmd: str) -> tuple[str, str] | None:
api_ops = (
"create-bucket",
"put-object",
"put-bucket-versioning",
"put-bucket-policy",
"get-bucket-policy",
"head-bucket",
)
for op in api_ops:
if f"aws s3api {op}" in cmd:
wrong = cmd.replace("aws s3api", "aws s3", 1)
return wrong, (
f"aws: error: argument operation: Invalid choice: '{op}'"
)
return None
def _mistake_typo_in_resource(cmd: str) -> tuple[str, str] | None:
m = re.search(
r"--(bucket|table-name|role-name|function-name|queue-name|name)\s+(\S+)", cmd
)
if not m:
return None
key, val = m.group(1), m.group(2)
if len(val) < 3:
return None
typo = val[0] + val[2] + val[1] + val[3:]
wrong = cmd.replace(f"--{key} {val}", f"--{key} {typo}", 1)
return wrong, (
f"An error occurred (NoSuchEntity): The resource '{typo}' was not found"
)
_MISTAKES: list[Callable[[str], tuple[str, str] | None]] = [
_mistake_wrong_operation,
_mistake_missing_arg,
_mistake_wrong_service,
_mistake_s3_vs_s3api,
_mistake_typo_in_resource,
]
def perturb_command(correct: str, rng: random.Random) -> tuple[str, str]:
order = list(_MISTAKES)
rng.shuffle(order)
for m in order:
out = m(correct)
if out is not None:
return out
return correct + " --foo bar", "aws: error: unknown option: --foo"
def _jitter_reward(base: float, rng: random.Random) -> float:
val = base + rng.uniform(-0.1, 0.1)
return max(0.0, min(1.0, round(val, 2)))
def _trim_history(history: list[str], rng: random.Random) -> list[str]:
if not history:
return history
options = [len(history), min(len(history), 4), min(len(history), 2)]
k = rng.choice(options)
return history[-k:]
def render_user_prompt(
task_description: str,
step: int,
last_output: str,
last_error: str,
last_reward: float,
history: list[str],
) -> str:
history_block = "\n".join(history) if history else "None"
return textwrap.dedent(
f"""
TASK: {task_description}
Step: {step}
Last command output: {last_output!r}
Last error: {last_error!r}
Last reward: {last_reward:.2f}
Previous steps:
{history_block}
Send your next AWS CLI command.
"""
).strip()
def make_row(
task: dict,
tier: str,
source: str,
step_idx: int,
user_prompt: str,
assistant_command: str,
) -> dict[str, Any]:
return {
"task_id": task["task_id"],
"difficulty": tier,
"source": source,
"step_idx": step_idx,
"messages": [
{"role": "system", "content": SYSTEM_PROMPT},
{"role": "user", "content": user_prompt},
{"role": "assistant", "content": assistant_command},
],
}
def produce_success_first_step(
task: dict, tier: str, commands: dict[int, list[str]], rng: random.Random
) -> dict | None:
cmds = commands.get(task["task_id"])
if not cmds:
return None
user = render_user_prompt(
task["description"],
step=0,
last_output=_sample_initial_output(rng),
last_error="",
last_reward=_jitter_reward(0.0, rng),
history=[],
)
return make_row(task, tier, "success_first_step", 0, user, cmds[0])
def produce_multi_step_continuation(
task: dict, tier: str, commands: dict[int, list[str]], rng: random.Random
) -> dict | None:
cmds = commands.get(task["task_id"])
if not cmds or len(cmds) < 2:
return None
i = rng.randint(1, len(cmds) - 1)
prior = cmds[:i]
last_output = simulate_output(prior[-1])
history = [f"{n + 1}. {c}" for n, c in enumerate(prior)]
history = _trim_history(history, rng)
base_reward = 0.2 + 0.6 * (i / len(cmds))
user = render_user_prompt(
task["description"],
step=i,
last_output=last_output,
last_error="",
last_reward=_jitter_reward(base_reward, rng),
history=history,
)
return make_row(task, tier, "multi_step_continuation", i, user, cmds[i])
def produce_failure_recovery(
task: dict, tier: str, commands: dict[int, list[str]], rng: random.Random
) -> dict | None:
cmds = commands.get(task["task_id"])
if not cmds:
return None
i = rng.randint(0, len(cmds) - 1)
correct = cmds[i]
wrong, err = perturb_command(correct, rng)
history = [f"{n + 1}. {c}" for n, c in enumerate(cmds[:i])]
history.append(f"{i + 1}. {wrong}")
history = _trim_history(history, rng)
step_now = i + 1
user = render_user_prompt(
task["description"],
step=step_now,
last_output="",
last_error=err,
last_reward=_jitter_reward(0.0 if i == 0 else 0.3, rng),
history=history,
)
return make_row(task, tier, "failure_recovery", step_now, user, correct)
_VERIFY_MAP: list[tuple[str, Callable[[str], str | None]]] = [
(
"aws s3api create-bucket",
lambda c: (
f"aws s3api head-bucket{_flag_passthrough(c, '--bucket')}"
if _flag_passthrough(c, "--bucket")
else None
),
),
(
"aws s3api put-object",
lambda c: (
f"aws s3api list-objects-v2{_flag_passthrough(c, '--bucket')}"
if _flag_passthrough(c, "--bucket")
else None
),
),
(
"aws s3api put-bucket-versioning",
lambda c: (
f"aws s3api get-bucket-versioning{_flag_passthrough(c, '--bucket')}"
if _flag_passthrough(c, "--bucket")
else None
),
),
(
"aws s3api put-bucket-policy",
lambda c: (
f"aws s3api get-bucket-policy{_flag_passthrough(c, '--bucket')}"
if _flag_passthrough(c, "--bucket")
else None
),
),
(
"aws dynamodb create-table",
lambda c: (
f"aws dynamodb describe-table{_flag_passthrough(c, '--table-name')}"
if _flag_passthrough(c, "--table-name")
else None
),
),
(
"aws dynamodb put-item",
lambda c: (
f"aws dynamodb scan{_flag_passthrough(c, '--table-name')}"
if _flag_passthrough(c, "--table-name")
else None
),
),
(
"aws iam create-role",
lambda c: (
f"aws iam get-role{_flag_passthrough(c, '--role-name')}"
if _flag_passthrough(c, "--role-name")
else None
),
),
(
"aws iam attach-role-policy",
lambda c: (
f"aws iam list-attached-role-policies{_flag_passthrough(c, '--role-name')}"
if _flag_passthrough(c, "--role-name")
else None
),
),
(
"aws iam create-policy",
lambda c: (
f"aws iam list-policies --scope Local"
if "--policy-name" in c
else None
),
),
(
"aws lambda create-function",
lambda c: (
f"aws lambda get-function{_flag_passthrough(c, '--function-name')}"
if _flag_passthrough(c, "--function-name")
else None
),
),
("aws sns create-topic", lambda c: "aws sns list-topics"),
("aws sqs create-queue", lambda c: "aws sqs list-queues"),
(
"aws secretsmanager create-secret",
lambda c: (
f"aws secretsmanager describe-secret{_flag_passthrough(c, '--name', flag_out='--secret-id')}"
if _flag_passthrough(c, "--name")
else None
),
),
]
def _flag_passthrough(cmd: str, flag: str, flag_out: str | None = None) -> str:
m = re.search(rf"{re.escape(flag)}\s+(\S+)", cmd)
if not m:
return ""
out_flag = flag_out or flag
return f" {out_flag} {m.group(1)}"
def produce_verification(
task: dict, tier: str, commands: dict[int, list[str]], rng: random.Random
) -> dict | None:
cmds = commands.get(task["task_id"])
if not cmds or len(cmds) < 2:
return None
last = cmds[-1]
verify: str | None = None
for prefix, fn in _VERIFY_MAP:
if last.startswith(prefix):
verify = fn(last)
if verify:
break
if not verify:
return None
history = [f"{n + 1}. {c}" for n, c in enumerate(cmds)]
history = _trim_history(history, rng)
last_output = simulate_output(cmds[-1])
step_now = len(cmds)
user = render_user_prompt(
task["description"],
step=step_now,
last_output=last_output,
last_error="",
last_reward=_jitter_reward(0.85, rng),
history=history,
)
return make_row(task, tier, "verification", step_now, user, verify)
def produce_hint_usage(
task: dict, tier: str, commands: dict[int, list[str]], rng: random.Random
) -> dict | None:
user = render_user_prompt(
task["description"],
step=0,
last_output=_sample_initial_output(rng),
last_error="",
last_reward=_jitter_reward(0.0, rng),
history=[],
)
return make_row(task, tier, "hint_usage", 0, user, "aws help --task-hint")
PRODUCERS: dict[
str, tuple[Callable[[dict, str, dict, random.Random], dict | None], list[str]]
] = {
"success_first_step": (
produce_success_first_step,
["warmup", "beginner", "intermediate", "advanced"],
),
"multi_step_continuation": (
produce_multi_step_continuation,
["intermediate", "advanced"],
),
"failure_recovery": (
produce_failure_recovery,
["warmup", "beginner", "intermediate", "advanced"],
),
"verification": (produce_verification, ["intermediate", "advanced"]),
"hint_usage": (produce_hint_usage, ["intermediate", "advanced"]),
}
def pick_tier(eligible_tiers: list[str], rng: random.Random) -> str | None:
weights = [TIER_WEIGHTS[t] for t in eligible_tiers]
if sum(weights) == 0:
return rng.choice(eligible_tiers) if eligible_tiers else None
return rng.choices(eligible_tiers, weights=weights, k=1)[0]
def build_dataset(
n: int,
tasks: dict[str, list[dict]],
commands: dict[int, list[str]],
rng: random.Random,
) -> list[dict]:
counts = {src: round(n * pct) for src, pct in SOURCE_MIX.items()}
diff = n - sum(counts.values())
counts["success_first_step"] += diff
rows: list[dict] = []
seen: set[tuple[str, str]] = set()
for source, want in counts.items():
producer, eligible_tiers = PRODUCERS[source]
made = 0
attempts = 0
max_attempts = max(want * 20, 100)
while made < want and attempts < max_attempts:
attempts += 1
tier = pick_tier(eligible_tiers, rng)
if tier is None:
break
tier_tasks = [
t
for t in tasks.get(tier, [])
if t["task_id"] in commands
and not task_has_dynamic_ids(commands[t["task_id"]])
]
if not tier_tasks:
continue
task = rng.choice(tier_tasks)
row = producer(task, tier, commands, rng)
if row is None:
continue
user_text = row["messages"][1]["content"]
asst_text = row["messages"][2]["content"]
key = (user_text, asst_text)
if key in seen:
continue
seen.add(key)
rows.append(row)
made += 1
if made < want:
print(
f" WARN: source '{source}' only produced {made}/{want} unique rows "
f"after {attempts} attempts"
)
rng.shuffle(rows)
return rows
def compute_stats(rows: list[dict]) -> dict:
if not rows:
return {"total": 0}
by_source = Counter(r["source"] for r in rows)
by_tier = Counter(r["difficulty"] for r in rows)
by_task = Counter(r["task_id"] for r in rows)
return {
"total": len(rows),
"by_source": dict(by_source),
"by_source_pct": {k: round(v / len(rows), 3) for k, v in by_source.items()},
"by_tier": dict(by_tier),
"by_tier_pct": {k: round(v / len(rows), 3) for k, v in by_tier.items()},
"unique_tasks": len(by_task),
"top_tasks": by_task.most_common(10),
}
def write_jsonl(rows: list[dict], path: Path) -> None:
path.parent.mkdir(parents=True, exist_ok=True)
with open(path, "w") as f:
for row in rows:
f.write(json.dumps(row, ensure_ascii=False) + "\n")
def main() -> None:
ap = argparse.ArgumentParser(description=__doc__.splitlines()[0])
ap.add_argument("--train", type=int, default=1500)
ap.add_argument("--val", type=int, default=150)
ap.add_argument("--reserve", type=int, default=200)
ap.add_argument("--seed", type=int, default=42)
ap.add_argument("--out-dir", type=Path, default=DEFAULT_OUT_DIR)
ap.add_argument(
"--repo-root",
type=Path,
default=None,
help="Override auto-detected repo root (must contain server/services/tasks/)",
)
args = ap.parse_args()
if args.repo_root is not None:
global REPO_ROOT, TASKS_DIR, TESTS_DIR
REPO_ROOT = args.repo_root.resolve()
TASKS_DIR = REPO_ROOT / "server" / "services" / "tasks"
TESTS_DIR = REPO_ROOT / "tests_tasks"
if not TASKS_DIR.is_dir():
raise SystemExit(
f"ERROR: task dir not found at {TASKS_DIR}\n"
f" Auto-detected repo root: {REPO_ROOT}\n"
f" Pass --repo-root <path-to-aws-rl-env> to override."
)
rng = random.Random(args.seed)
tasks = load_tasks()
commands = load_canonical_commands()
task_count = sum(len(v) for v in tasks.values())
print(f"Loaded {task_count} tasks across {len(tasks)} tiers:")
for tier in TIERS:
tier_tasks = tasks.get(tier, [])
with_cmd = sum(1 for t in tier_tasks if t["task_id"] in commands)
with_cmd_no_dyn = sum(
1
for t in tier_tasks
if t["task_id"] in commands
and not task_has_dynamic_ids(commands[t["task_id"]])
)
print(
f" {tier:<13} {len(tier_tasks):3d} tasks "
f"({with_cmd} w/ canonical cmds, {with_cmd_no_dyn} after dynamic-id filter)"
)
print(f"Canonical commands loaded for {len(commands)} task IDs\n")
total = args.train + args.val + args.reserve
print(f"Building {total} rows (train={args.train} val={args.val} reserve={args.reserve})")
all_rows = build_dataset(total, tasks, commands, rng)
if len(all_rows) < total:
print(f"\n WARN: only built {len(all_rows)}/{total} unique rows")
print(" Splits will be proportionally shrunk to preserve train/val/reserve ratio.\n")
ratio = len(all_rows) / total
train_n = int(args.train * ratio)
val_n = int(args.val * ratio)
reserve_n = len(all_rows) - train_n - val_n
else:
train_n, val_n, reserve_n = args.train, args.val, args.reserve
train_rows = all_rows[:train_n]
val_rows = all_rows[train_n : train_n + val_n]
reserve_rows = all_rows[train_n + val_n : train_n + val_n + reserve_n]
out_dir = args.out_dir
write_jsonl(train_rows, out_dir / "aws_rl_sft.train.jsonl")
write_jsonl(val_rows, out_dir / "aws_rl_sft.val.jsonl")
write_jsonl(reserve_rows, out_dir / "aws_rl_sft.reserve.jsonl")
stats = {
"train": compute_stats(train_rows),
"val": compute_stats(val_rows),
"reserve": compute_stats(reserve_rows),
"targets": {"source_mix": SOURCE_MIX, "tier_weights": TIER_WEIGHTS},
"seed": args.seed,
}
with open(out_dir / "dataset_stats.json", "w") as f:
json.dump(stats, f, indent=2)
print(f"\nWrote:")
print(f" {out_dir / 'aws_rl_sft.train.jsonl'} ({len(train_rows)} rows)")
print(f" {out_dir / 'aws_rl_sft.val.jsonl'} ({len(val_rows)} rows)")
print(f" {out_dir / 'aws_rl_sft.reserve.jsonl'} ({len(reserve_rows)} rows)")
print(f" {out_dir / 'dataset_stats.json'}")
print(f"\nTrain source mix: {stats['train'].get('by_source_pct', {})}")
print(f"Train tier mix: {stats['train'].get('by_tier_pct', {})}")
if __name__ == "__main__":
main()
|