Spaces:
Running on Zero
Running on Zero
File size: 16,659 Bytes
0dd6c2f | 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 | from __future__ import annotations
import math
import random
from collections.abc import Sequence
from linalg_zero.grpo.types import CurriculumConfig, Task
class ShuffleBagSampler:
"""
Coverage-guaranteeing sampler ("shuffle bag").
Repeated calls to `sample_batch()` return indices from the eligible pool such that:
- No index repeats until all currently-eligible indices have been returned once (a "cycle"),
unless `batch_size` exceeds the pool size.
- If the eligible pool grows over time (curriculum), newly added indices are injected into the
current cycle so they are surfaced before any repeats.
This is deterministic given the same `seed` and the same sequence of eligible pools.
"""
def __init__(self, *, seed: int = 0, shuffle: bool = True) -> None:
self._rng = random.Random(int(seed))
self._shuffle = bool(shuffle)
self._pool_set: set[int] = set()
self._pool_order: list[int] = []
self._seen_in_cycle: set[int] = set()
self._remaining: list[int] = []
self._remaining_set: set[int] = set()
def _reset_cycle(self) -> None:
self._seen_in_cycle.clear()
self._remaining = list(self._pool_order)
if self._shuffle and len(self._remaining) > 1:
self._rng.shuffle(self._remaining)
self._remaining_set = set(self._remaining)
def _update_pool(self, *, eligible: Sequence[int]) -> None:
eligible_list = list(eligible)
eligible_set = set(eligible_list)
if not eligible_set:
raise ValueError("Cannot sample from an empty index set.")
if not self._pool_set:
self._pool_set = eligible_set
self._pool_order = eligible_list
self._reset_cycle()
return
added = [idx for idx in eligible_list if idx not in self._pool_set]
removed = self._pool_set - eligible_set
if not added and not removed:
return
if removed:
self._pool_set.difference_update(removed)
self._seen_in_cycle.difference_update(removed)
if self._remaining:
self._remaining = [idx for idx in self._remaining if idx not in removed]
self._remaining_set.difference_update(removed)
if added:
self._pool_set.update(added)
# Inject newly-eligible tasks into the current cycle so we don't defer them until the next cycle.
for idx in added:
if idx in self._seen_in_cycle or idx in self._remaining_set:
continue
pos = self._rng.randrange(len(self._remaining) + 1)
self._remaining.insert(pos, idx)
self._remaining_set.add(idx)
self._pool_order = eligible_list
# If we removed indices, it's possible to end up with an empty remaining set mid-cycle.
# Start a fresh cycle in that case.
if not self._remaining:
self._reset_cycle()
def sample_batch(self, *, eligible: Sequence[int], batch_size: int) -> list[int]:
"""
Sample a batch of indices with coverage guarantees.
If `batch_size > len(eligible)`, repeats are unavoidable; this will still maximize coverage by
cycling through the pool as many times as needed.
"""
if batch_size <= 0:
return []
self._update_pool(eligible=eligible)
out: list[int] = []
while len(out) < batch_size:
if not self._remaining:
self._reset_cycle()
idx = self._remaining.pop()
self._remaining_set.remove(idx)
self._seen_in_cycle.add(idx)
out.append(idx)
return out
def _clamp01(value: float) -> float:
return float(max(0.0, min(1.0, value)))
def _deterministic_counts_from_probs(*, probs: Sequence[float], total: int) -> list[int]:
if total <= 0:
return [0 for _ in probs]
cleaned: list[float] = [max(0.0, float(p)) for p in probs]
mass = sum(cleaned)
if mass <= 0.0:
# Fall back to uniform allocation if the distribution degenerates.
cleaned = [1.0 for _ in probs]
mass = float(len(cleaned))
normed = [p / mass for p in cleaned]
expected = [p * total for p in normed]
# Deterministic rounding: first take the floor of each expected count.
counts = [math.floor(e) for e in expected]
remainder = total - sum(counts)
if remainder <= 0:
return counts
# Then allocate the leftover `remainder` to buckets with the largest fractional remainders
# (`expected - floor(expected)`), not to buckets with the largest already-integer `counts`.
fractional = [(expected[i] - counts[i], i) for i in range(len(counts))]
fractional.sort(key=lambda item: (-item[0], item[1]))
for _, idx in fractional[:remainder]:
counts[idx] += 1
return counts
class ToolCallsMixtureSampler:
"""
Deterministic per-step mixture sampler over tool-call "difficulty" buckets.
- Buckets are defined by `len(task.actions)` (teacher tool calls).
- Each bucket uses its own ShuffleBagSampler to maximize coverage within the bucket.
- Per-step mixture weights are computed from a Gaussian centered on a target tool-call
count that increases linearly with `difficulty`.
"""
def __init__(
self,
*,
tasks: Sequence[Task],
indices: Sequence[int],
curriculum: CurriculumConfig,
seed: int = 0,
shuffle: bool = True,
) -> None:
if curriculum.metric != "tool_calls":
raise ValueError(f"Unknown curriculum metric: {curriculum.metric!r}")
self._tasks = tasks
self._curriculum = curriculum
self._seed = int(seed)
self._shuffle = bool(shuffle)
self._calls = 0
base_indices = list(indices)
if not base_indices:
raise ValueError("Cannot create curriculum sampler from an empty index set.")
buckets: dict[int, list[int]] = {}
for idx in base_indices:
tool_calls = len(tasks[idx].actions)
buckets.setdefault(tool_calls, []).append(idx)
self._min_tool_calls = min(buckets.keys())
max_tool_calls_seen = max(buckets.keys())
initial_tool_calls = max(self._min_tool_calls, int(curriculum.initial_max_tool_calls))
requested_final = (
max_tool_calls_seen
if curriculum.final_max_tool_calls is None
else min(int(curriculum.final_max_tool_calls), max_tool_calls_seen)
)
requested_final = max(self._min_tool_calls, requested_final)
if requested_final < initial_tool_calls:
raise ValueError(
"Invalid curriculum: final_max_tool_calls must be >= initial_max_tool_calls "
f"(got initial={initial_tool_calls}, final={requested_final})."
)
self._final_tool_calls = requested_final
filtered_keys = sorted([k for k in buckets if k <= self._final_tool_calls])
if not filtered_keys:
raise ValueError("Curriculum sampler has no buckets after applying final_max_tool_calls filter.")
self._bucket_keys = filtered_keys
self._bucket_ordered: dict[int, list[int]] = {}
self._bucket_samplers: dict[int, ShuffleBagSampler] = {}
for tool_calls in self._bucket_keys:
bucket = list(buckets[tool_calls])
if self._shuffle and len(bucket) > 1:
rng = random.Random(self._seed + (tool_calls + 1) * 1_000_003)
rng.shuffle(bucket)
self._bucket_ordered[tool_calls] = bucket
self._bucket_samplers[tool_calls] = ShuffleBagSampler(
seed=self._seed + (tool_calls + 1) * 2_000_033,
shuffle=self._shuffle,
)
def _target_tool_calls(self, *, difficulty: float) -> float:
difficulty = _clamp01(difficulty)
start = float(max(self._min_tool_calls, int(self._curriculum.initial_max_tool_calls)))
end = float(max(start, self._final_tool_calls))
return float(max(self._min_tool_calls, min(end, start + difficulty * (end - start))))
def _mixture_probs(self, *, difficulty: float) -> list[float]:
target = self._target_tool_calls(difficulty=difficulty)
sigma = float(getattr(self._curriculum, "mixture_sigma", 0.0))
if sigma <= 0.0:
# Hard assignment to the nearest bucket when sigma is disabled.
nearest = min(self._bucket_keys, key=lambda tc: (abs(float(tc) - target), tc))
return [1.0 if tc == nearest else 0.0 for tc in self._bucket_keys]
denom = 2.0 * sigma * sigma
if denom == 0.0:
# Extremely small positive sigma can underflow `sigma * sigma` to 0.0.
nearest = min(self._bucket_keys, key=lambda tc: (abs(float(tc) - target), tc))
return [1.0 if tc == nearest else 0.0 for tc in self._bucket_keys]
log_weights = [-((float(tc) - target) ** 2) / denom for tc in self._bucket_keys]
max_log = max(log_weights)
weights = [math.exp(w - max_log) for w in log_weights]
mass = sum(weights)
if mass <= 0.0:
# Extremely small sigma can cause underflow; fall back to nearest-bucket.
nearest = min(self._bucket_keys, key=lambda tc: (abs(float(tc) - target), tc))
probs = [1.0 if tc == nearest else 0.0 for tc in self._bucket_keys]
else:
probs = [w / mass for w in weights]
floor = float(getattr(self._curriculum, "mixture_min_prob_easiest", 0.0))
floor = float(max(0.0, min(1.0, floor)))
if floor <= 0.0:
return probs
p_easy = probs[0]
if p_easy >= floor or p_easy >= 1.0:
return probs
remaining = 1.0 - p_easy
if remaining <= 0.0:
return [1.0] + [0.0 for _ in probs[1:]]
scale = (1.0 - floor) / remaining
adjusted = [floor] + [p * scale for p in probs[1:]]
return adjusted
def sample_batch(self, *, difficulty: float, batch_size: int) -> list[int]:
if batch_size <= 0:
return []
difficulty = _clamp01(difficulty)
probs = self._mixture_probs(difficulty=difficulty)
# Convert continuous mixture probabilities into an exact integer allocation for this step.
# Note: when a bucket's expected count is <1 (e.g., early 3-tool-call exposure with small `p3`),
# deterministic rounding can "flicker" between 0 and 1 across steps; once the expected count
# crosses 1 (as difficulty increases), that bucket will appear every step thereafter.
counts = _deterministic_counts_from_probs(probs=probs, total=batch_size)
frac = self._curriculum.fraction_at_start + difficulty * (
self._curriculum.fraction_at_end - self._curriculum.fraction_at_start
)
frac = _clamp01(frac)
out: list[int] = []
for tool_calls, count in zip(self._bucket_keys, counts, strict=True):
if count <= 0:
continue
ordered = self._bucket_ordered[tool_calls]
k = math.ceil(frac * len(ordered))
if k <= 0:
k = 1
eligible = ordered[:k]
sampler = self._bucket_samplers[tool_calls]
out.extend(sampler.sample_batch(eligible=eligible, batch_size=count))
if len(out) != batch_size:
raise RuntimeError(f"Mixture sampler produced {len(out)} indices, expected {batch_size}.")
# Mix difficulties within the step so the batch isn't ordered by bucket.
# Deterministic given seed and call order (which is driven by the training step).
rng = random.Random(self._seed + 9_000_001 + self._calls)
rng.shuffle(out)
self._calls += 1
return out
def get_task_indices( # noqa: C901
*,
task_ids: list[int] | None,
start_index: int,
end_index: int,
tasks_length: int | None = None,
tasks: Sequence[Task] | None = None,
curriculum: CurriculumConfig | None = None,
difficulty: float | None = None,
seed: int = 0,
) -> list[int]:
"""
Return a list of task indices, optionally filtered by a curriculum.
- If `task_ids` is provided, it always wins (curriculum is ignored).
- Otherwise uses `[start_index, end_index)` (or full length if `end_index == -1`).
- If `curriculum.enabled` and `difficulty` is provided, returns a deterministic subset
that grows monotonically with `difficulty` (0.0 -> easiest subset, 1.0 -> full set).
"""
if task_ids:
return list(task_ids)
if tasks_length is None:
if tasks is None:
raise ValueError("Must provide `tasks_length` or `tasks` when task_ids is not set.")
tasks_length = len(tasks)
actual_start = max(0, start_index)
actual_end = tasks_length if end_index == -1 else min(end_index, tasks_length)
base_indices = list(range(actual_start, max(actual_start, actual_end)))
if curriculum is None or not curriculum.enabled or difficulty is None:
return base_indices
if tasks is None:
raise ValueError("Curriculum selection requires `tasks` to be provided.")
if curriculum.metric != "tool_calls":
raise ValueError(f"Unknown curriculum metric: {curriculum.metric!r}")
difficulty = float(max(0.0, min(1.0, difficulty)))
tool_calls_by_index: dict[int, int] = {}
max_tool_calls_seen = 0
min_tool_calls_seen: int | None = None
for idx in base_indices:
tool_calls = len(tasks[idx].actions)
tool_calls_by_index[idx] = tool_calls
max_tool_calls_seen = max(max_tool_calls_seen, tool_calls)
min_tool_calls_seen = tool_calls if min_tool_calls_seen is None else min(min_tool_calls_seen, tool_calls)
initial_max = max(0, int(curriculum.initial_max_tool_calls))
final_max = (
max_tool_calls_seen if curriculum.final_max_tool_calls is None else int(curriculum.final_max_tool_calls)
)
if final_max < initial_max:
final_max = initial_max
levels = max(1, final_max - initial_max + 1)
level = min(levels - 1, math.floor(difficulty * levels))
allowed_max = initial_max + level
if min_tool_calls_seen is not None:
allowed_max = max(allowed_max, min_tool_calls_seen)
# Fraction of tasks to expose within each included difficulty bucket.
frac = curriculum.fraction_at_start + difficulty * (curriculum.fraction_at_end - curriculum.fraction_at_start)
frac = float(max(0.0, min(1.0, frac)))
buckets: dict[int, list[int]] = {}
for idx in base_indices:
tool_calls = tool_calls_by_index[idx]
if tool_calls <= allowed_max:
buckets.setdefault(tool_calls, []).append(idx)
ordered_buckets: dict[int, list[int]] = {}
for tool_calls in sorted(buckets.keys()):
bucket = list(buckets[tool_calls])
rng = random.Random(seed + (tool_calls + 1) * 1_000_003)
rng.shuffle(bucket)
ordered_buckets[tool_calls] = bucket
# Stable "easiest-first" ordering so that when new buckets are unlocked,
# previously selected tasks remain selected.
easiest_first: list[int] = []
for tool_calls in sorted(ordered_buckets.keys()):
easiest_first.extend(ordered_buckets[tool_calls])
# Baseline floor (picked from easiest-first) to keep selection non-empty and monotonic.
min_total = max(0, int(curriculum.min_total_tasks))
baseline = set(easiest_first[: min(min_total, len(easiest_first))])
# Fraction-based selection (per bucket), chosen as prefixes of deterministic per-bucket order.
fraction_selected: set[int] = set()
for tool_calls in sorted(ordered_buckets.keys()):
bucket = ordered_buckets[tool_calls]
k = math.ceil(frac * len(bucket))
if k <= 0:
continue
fraction_selected.update(bucket[:k])
selected_set = baseline | fraction_selected
if not selected_set and easiest_first:
selected_set = {easiest_first[0]}
return [idx for idx in easiest_first if idx in selected_set]
def sample_indices_to_length(*, indices: Sequence[int], length: int, rng: random.Random) -> list[int]:
"""Sample `length` indices, using replacement if needed."""
if length <= 0:
return []
if not indices:
raise ValueError("Cannot sample from an empty index set.")
if len(indices) >= length:
return list(rng.sample(list(indices), length))
return list(rng.choices(list(indices), k=length))
|