Spaces:
Running
Running
File size: 19,221 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 | """Curriculum manager for progressive LLM training in the AWS RL environment.
Training flow:
1. Agent starts at the warmup tier with simple listing tasks.
2. A priority queue selects the next task based on weakness, novelty,
spaced repetition, and recency β replacing blind round-robin.
3. Per-task mastery tracking graduates individual tasks once the agent
demonstrates sustained competence.
4. Graduated tasks resurface via spaced repetition at exponentially
increasing intervals to prevent catastrophic forgetting.
5. Fast-track promotion lets strong agents skip minimum episode waits.
6. Exponential decay on history ensures recent results matter more.
"""
import heapq
import logging
import random
from collections import defaultdict
from pathlib import Path
from typing import Any
import yaml
from models import (
SetupCommand,
SpacedRepState,
SuccessCriteria,
Task,
TaskDifficulty,
TaskID,
TierConfig,
)
logger = logging.getLogger(__name__)
TASKS_DIR = Path(__file__).parent / "tasks"
# ---------------------------------------------------------------------------
# Per-tier configuration
# ---------------------------------------------------------------------------
TIER_CONFIGS: dict[TaskDifficulty, TierConfig] = {
TaskDifficulty.WARMUP: TierConfig(
min_episodes=5,
advance_rate=0.6,
mastery_window=10,
mastery_threshold=0.7,
fast_track_rate=0.9,
),
TaskDifficulty.BEGINNER: TierConfig(
min_episodes=5,
advance_rate=0.6,
mastery_window=10,
mastery_threshold=0.7,
fast_track_rate=0.9,
),
TaskDifficulty.INTERMEDIATE: TierConfig(
min_episodes=8,
advance_rate=0.65,
mastery_window=10,
mastery_threshold=0.7,
fast_track_rate=0.9,
chaos_probability=0.1,
),
TaskDifficulty.ADVANCED: TierConfig(
min_episodes=10,
advance_rate=0.7,
mastery_window=10,
mastery_threshold=0.7,
fast_track_rate=0.9,
chaos_probability=0.2,
),
TaskDifficulty.EXPERT: TierConfig(
min_episodes=0,
advance_rate=1.0,
mastery_window=10,
mastery_threshold=0.7,
fast_track_rate=1.0,
chaos_probability=0.3,
),
}
# Map YAML filenames to difficulty tiers
_TIER_FILES: dict[TaskDifficulty, str] = {
TaskDifficulty.WARMUP: "warmup.yaml",
TaskDifficulty.BEGINNER: "beginner.yaml",
TaskDifficulty.INTERMEDIATE: "intermediate.yaml",
TaskDifficulty.ADVANCED: "advanced.yaml",
TaskDifficulty.EXPERT: "expert.yaml",
}
# Supplementary task files merged into an existing tier
_SUPPLEMENTARY_FILES: dict[TaskDifficulty, list[str]] = {
TaskDifficulty.EXPERT: ["drift.yaml"],
}
# ---------------------------------------------------------------------------
# Priority score tuning constants
# ---------------------------------------------------------------------------
_NOVELTY_BONUS = 100 # untried tasks β explore first
_WEAKNESS_WEIGHT = 50 # multiplied by (1 - success_rate)
_SPACED_REP_BONUS = 30 # graduated task due for re-test
_RECENCY_PENALTY = 20 # attempted in last 2 episodes
# Exponential decay factor for weighted success rate
_DECAY_FACTOR = 0.85
# Minimum attempts before a task can be graduated
_MIN_ATTEMPTS_FOR_MASTERY = 3
# Fast-track requires at least this many episodes in the tier
_FAST_TRACK_MIN_EPISODES = 3
# ---------------------------------------------------------------------------
# YAML loader
# ---------------------------------------------------------------------------
def _parse_task_entries(
entries: list[dict[str, Any]], difficulty: TaskDifficulty
) -> list[Task]:
"""Convert raw YAML entries into Task models."""
return [
Task(
task_id=TaskID(entry["task_id"]),
difficulty=difficulty,
description=entry["description"],
success_criteria=SuccessCriteria(**entry.get("success_criteria", {})),
setup_commands=[
SetupCommand(command=cmd)
if isinstance(cmd, str)
else SetupCommand(**cmd)
for cmd in entry.get("setup_commands", [])
],
desired_state_spec=entry.get("desired_state_spec"),
possible_drifts=[
SetupCommand(command=d) if isinstance(d, str) else SetupCommand(**d)
for d in entry.get("possible_drifts", [])
],
)
for entry in entries
]
def load_tier(difficulty: TaskDifficulty, tasks_dir: Path = TASKS_DIR) -> list[Task]:
"""Load tasks for a single difficulty tier from its YAML file(s)."""
filename = _TIER_FILES.get(difficulty)
if filename is None:
logger.warning("No file mapping for difficulty: %s", difficulty.value)
return []
filepath = tasks_dir / filename
if not filepath.exists():
logger.warning("Task file not found: %s", filepath)
return []
with open(filepath) as f:
entries = yaml.safe_load(f) or []
tasks = _parse_task_entries(entries, difficulty)
# Load supplementary task files for this tier
for extra_file in _SUPPLEMENTARY_FILES.get(difficulty, []):
extra_path = tasks_dir / extra_file
if not extra_path.exists():
continue
with open(extra_path) as f:
extra_entries = yaml.safe_load(f) or []
extra_tasks = _parse_task_entries(extra_entries, difficulty)
tasks.extend(extra_tasks)
logger.info(
"Loaded %d supplementary %s tasks from %s",
len(extra_tasks),
difficulty.value,
extra_file,
)
logger.info("Loaded %d %s tasks total", len(tasks), difficulty.value)
return tasks
# ---------------------------------------------------------------------------
# Helpers
# ---------------------------------------------------------------------------
def _weighted_success_rate(results: list[bool], decay: float = _DECAY_FACTOR) -> float:
"""Compute success rate with exponential decay β recent results matter more."""
if not results:
return 0.0
weights = [decay**i for i in range(len(results) - 1, -1, -1)]
total_weight = sum(weights)
return sum(w * float(r) for w, r in zip(weights, results)) / total_weight
# ---------------------------------------------------------------------------
# Curriculum
# ---------------------------------------------------------------------------
class Curriculum:
"""Manages progressive task assignment with priority-queue-based selection.
Features:
- Priority queue task selection (novelty, weakness, spaced rep, recency)
- Per-task mastery tracking with graduation
- Spaced repetition for graduated tasks (prevents catastrophic forgetting)
- Fast-track tier promotion for strong agents
- Exponential decay on success history
- Rich observability via get_stats()
"""
def __init__(
self,
tier_configs: dict[TaskDifficulty, TierConfig] | None = None,
tasks_dir: Path = TASKS_DIR,
) -> None:
self._tier_configs = tier_configs or TIER_CONFIGS
self._tasks_dir = tasks_dir
# Ordered difficulty progression
self._levels = list(TaskDifficulty)
# Tier tracking
self._current_level_idx: int = 0
self._tier_episodes: int = 0
self._tier_results: list[bool] = [] # results within current tier
# Per-task tracking
self._task_history: dict[TaskID, list[bool]] = defaultdict(list)
self._task_attempt_count: dict[TaskID, int] = defaultdict(int)
self._last_attempted_episode: dict[TaskID, int] = {}
self._graduated_tasks: set[TaskID] = set()
self._spaced_rep: dict[TaskID, SpacedRepState] = {}
# Global counters
self._episode_count: int = 0
self._episode_rewards: list[float] = []
# Load starting tier
self._current_tasks: list[Task] = load_tier(
self.current_difficulty, self._tasks_dir
)
self._task_map: dict[TaskID, Task] = {t.task_id: t for t in self._current_tasks}
# Priority queue: list of (-score, random_tiebreaker, task_id)
self._priority_queue: list[tuple[float, float, TaskID]] = []
self._rebuild_priority_queue()
logger.info(
"Curriculum initialised β starting at %s with %d tasks",
self.current_difficulty.value,
len(self._current_tasks),
)
# -- Properties -----------------------------------------------------------
@property
def current_difficulty(self) -> TaskDifficulty:
return self._levels[self._current_level_idx]
@property
def tier_config(self) -> TierConfig:
return self._tier_configs[self.current_difficulty]
@property
def current_level_success_rate(self) -> float:
return _weighted_success_rate(self._tier_results)
@property
def is_warmup(self) -> bool:
return self.current_difficulty == TaskDifficulty.WARMUP
@property
def chaos_probability(self) -> float:
return self.tier_config.chaos_probability
# -- Public API -----------------------------------------------------------
def next_task(self) -> Task:
"""Select the highest-priority task from the current tier."""
if not self._current_tasks:
self._current_tasks = load_tier(self.current_difficulty, self._tasks_dir)
self._task_map = {t.task_id: t for t in self._current_tasks}
self._rebuild_priority_queue()
if not self._priority_queue:
self._rebuild_priority_queue()
# Pop highest priority (most negative = highest score)
_, _, task_id = heapq.heappop(self._priority_queue)
task = self._task_map[task_id]
# If queue is now empty, rebuild for next call
if not self._priority_queue:
self._rebuild_priority_queue()
return task
def get_task_by_id(self, task_id: TaskID) -> Task:
"""Look up a task by id, searching across all tiers if needed.
Used by GRPO training to force all rollouts in a group onto the same
task, bypassing the per-env priority queue.
"""
if task_id in self._task_map:
return self._task_map[task_id]
for difficulty in self._levels:
for task in load_tier(difficulty, self._tasks_dir):
if task.task_id == task_id:
return task
raise KeyError(f"task_id={task_id} not found in any tier")
def record_result(self, task: Task, achieved: bool, reward: float = 0.0) -> None:
"""Record episode outcome, update mastery, check promotion."""
self._episode_count += 1
self._tier_episodes += 1
self._episode_rewards.append(reward)
# Per-tier results
self._tier_results.append(achieved)
# Per-task results
self._task_history[task.task_id].append(achieved)
self._task_attempt_count[task.task_id] += 1
self._last_attempted_episode[task.task_id] = self._episode_count
# Check mastery
self._check_mastery(task.task_id)
# Check tier promotion
self._maybe_promote()
# Rebuild priority queue with updated scores
self._rebuild_priority_queue()
logger.info(
"Episode %d: task=%d difficulty=%s achieved=%s tier_rate=%.2f",
self._episode_count,
task.task_id,
task.difficulty.value,
achieved,
self.current_level_success_rate,
)
def reset(self) -> None:
"""Reset curriculum back to warmup (full training restart)."""
self._current_level_idx = 0
self._tier_episodes = 0
self._tier_results.clear()
self._task_history.clear()
self._task_attempt_count.clear()
self._last_attempted_episode.clear()
self._graduated_tasks.clear()
self._spaced_rep.clear()
self._episode_count = 0
self._episode_rewards.clear()
self._current_tasks = load_tier(self.current_difficulty, self._tasks_dir)
self._task_map = {t.task_id: t for t in self._current_tasks}
self._rebuild_priority_queue()
logger.info("Curriculum reset to %s", self.current_difficulty.value)
# -- Observability --------------------------------------------------------
def get_skill_profile(self) -> dict[TaskID, float]:
"""Weighted success rate per task over recent history."""
config = self.tier_config
return {
task_id: round(_weighted_success_rate(results[-config.mastery_window :]), 2)
for task_id, results in self._task_history.items()
if results
}
def get_weak_spots(self) -> list[TaskID]:
"""Tasks in the current tier below mastery threshold."""
config = self.tier_config
profile = self.get_skill_profile()
return [
task_id
for task_id in self._task_map
if profile.get(task_id, 0.0) < config.mastery_threshold
and task_id not in self._graduated_tasks
]
def get_stats(self) -> dict:
"""Full curriculum state for logging/debugging."""
return {
"episode_count": self._episode_count,
"tier": self.current_difficulty.value,
"tier_episodes": self._tier_episodes,
"tier_success_rate": round(self.current_level_success_rate, 3),
"graduated_tasks": sorted(self._graduated_tasks),
"weak_spots": self.get_weak_spots(),
"skill_profile": self.get_skill_profile(),
"spaced_rep_due": [
int(tid) for tid in self._task_map if self._is_spaced_rep_due(tid)
],
"avg_reward_last_10": round(
sum(self._episode_rewards[-10:])
/ max(1, len(self._episode_rewards[-10:])),
3,
),
}
# -- Priority queue -------------------------------------------------------
def _compute_priority(self, task_id: TaskID) -> float:
"""Compute composite priority score for a task. Higher = selected sooner."""
config = self.tier_config
score = 0.0
attempts = self._task_attempt_count.get(task_id, 0)
# Novelty: never attempted β explore first
if attempts == 0:
score += _NOVELTY_BONUS
return score # no other signals available yet
# Weakness: worse tasks get higher priority
results = self._task_history.get(task_id, [])
task_rate = _weighted_success_rate(results[-config.mastery_window :])
score += _WEAKNESS_WEIGHT * (1.0 - task_rate)
# Spaced repetition: graduated task due for re-test
if task_id in self._graduated_tasks and self._is_spaced_rep_due(task_id):
score += _SPACED_REP_BONUS
# Recency penalty: attempted in last 2 episodes
last_ep = self._last_attempted_episode.get(task_id, -100)
if self._episode_count - last_ep <= 2:
score -= _RECENCY_PENALTY
return score
def _rebuild_priority_queue(self) -> None:
"""Recompute priorities for all current-tier tasks and rebuild the heap."""
self._priority_queue.clear()
for task in self._current_tasks:
score = self._compute_priority(task.task_id)
# heapq is a min-heap, so negate score for max-priority-first
# random tiebreaker prevents deterministic ordering among equal scores
heapq.heappush(
self._priority_queue,
(-score, random.random(), task.task_id),
)
# -- Mastery & spaced repetition ------------------------------------------
def _check_mastery(self, task_id: TaskID) -> None:
"""Check if a task should be graduated or un-graduated."""
config = self.tier_config
results = self._task_history.get(task_id, [])
recent = results[-config.mastery_window :]
if len(recent) < _MIN_ATTEMPTS_FOR_MASTERY:
return
rate = _weighted_success_rate(recent)
if rate >= config.mastery_threshold:
if task_id not in self._graduated_tasks:
self._graduated_tasks.add(task_id)
self._spaced_rep[task_id] = SpacedRepState(
interval=3,
last_graduated_episode=self._episode_count,
)
logger.info(
"Task %d GRADUATED (rate=%.2f) β scheduling spaced repetition",
task_id,
rate,
)
else:
# Un-graduate if performance dropped
if task_id in self._graduated_tasks:
self._graduated_tasks.discard(task_id)
self._spaced_rep.pop(task_id, None)
logger.info(
"Task %d UN-GRADUATED (rate=%.2f) β resetting to active",
task_id,
rate,
)
def _is_spaced_rep_due(self, task_id: TaskID) -> bool:
"""Check if a graduated task is due for a re-test."""
state = self._spaced_rep.get(task_id)
if state is None:
return False
episodes_since = self._episode_count - state.last_graduated_episode
return episodes_since >= state.interval
def _advance_spaced_rep(self, task_id: TaskID) -> None:
"""Double the interval after a successful re-test."""
state = self._spaced_rep.get(task_id)
if state is not None:
state.interval = min(state.interval * 2, 48) # cap at 48 episodes
state.last_graduated_episode = self._episode_count
# -- Tier promotion -------------------------------------------------------
def _maybe_promote(self) -> None:
"""Advance to the next difficulty tier if the agent is ready."""
if self._current_level_idx >= len(self._levels) - 1:
return # already at max tier
config = self.tier_config
rate = self.current_level_success_rate
# Fast-track: high success rate after minimum 3 episodes
fast_track = (
self._tier_episodes >= _FAST_TRACK_MIN_EPISODES
and rate >= config.fast_track_rate
)
if not fast_track and self._tier_episodes < config.min_episodes:
return
if rate < config.advance_rate:
return
prev_tier = self.current_difficulty.value
prev_rate = rate
self._current_level_idx += 1
self._tier_episodes = 0
self._tier_results.clear()
self._current_tasks = load_tier(self.current_difficulty, self._tasks_dir)
self._task_map = {t.task_id: t for t in self._current_tasks}
self._rebuild_priority_queue()
logger.info(
"PROMOTED from %s to %s (rate=%.2f%s)",
prev_tier,
self.current_difficulty.value,
prev_rate,
", FAST-TRACK" if fast_track else "",
)
|