Spaces:
Running
Running
File size: 15,707 Bytes
40ee6b4 |
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 |
"""
Data Splitting Module for Training Pipeline.
Provides utilities for:
- Train/validation/test splitting
- Stratified sampling by domain or difficulty
- Cross-validation fold creation
- Reproducible splits with seeding
"""
import logging
from collections import defaultdict
from dataclasses import dataclass
from typing import Any
from .dataset_loader import DatasetSample
logger = logging.getLogger(__name__)
@dataclass
class DataSplit:
"""Result of dataset splitting."""
train: list[DatasetSample]
validation: list[DatasetSample]
test: list[DatasetSample]
split_info: dict[str, Any]
@dataclass
class CrossValidationFold:
"""Single fold for cross-validation."""
fold_id: int
train: list[DatasetSample]
validation: list[DatasetSample]
class DataSplitter:
"""
Basic dataset splitter with random sampling.
Provides reproducible train/validation/test splits
with configurable ratios.
"""
def __init__(self, seed: int = 42):
"""
Initialize splitter.
Args:
seed: Random seed for reproducibility
"""
self.seed = seed
import random
self.rng = random.Random(seed)
def split(
self,
samples: list[DatasetSample],
train_ratio: float = 0.7,
val_ratio: float = 0.15,
test_ratio: float = 0.15,
shuffle: bool = True,
) -> DataSplit:
"""
Split dataset into train/validation/test sets.
Args:
samples: List of all samples
train_ratio: Proportion for training (default 0.7)
val_ratio: Proportion for validation (default 0.15)
test_ratio: Proportion for testing (default 0.15)
shuffle: Whether to shuffle before splitting
Returns:
DataSplit with train, validation, and test sets
"""
if abs(train_ratio + val_ratio + test_ratio - 1.0) > 0.001:
raise ValueError("Ratios must sum to 1.0")
if not samples:
raise ValueError("Cannot split empty sample list")
# Copy and optionally shuffle
all_samples = list(samples)
if shuffle:
self.rng.shuffle(all_samples)
n = len(all_samples)
train_end = int(n * train_ratio)
val_end = train_end + int(n * val_ratio)
train_samples = all_samples[:train_end]
val_samples = all_samples[train_end:val_end]
test_samples = all_samples[val_end:]
split_info = {
"total_samples": n,
"train_samples": len(train_samples),
"val_samples": len(val_samples),
"test_samples": len(test_samples),
"train_ratio": len(train_samples) / n,
"val_ratio": len(val_samples) / n,
"test_ratio": len(test_samples) / n,
"seed": self.seed,
"shuffled": shuffle,
}
logger.info(f"Split {n} samples: train={len(train_samples)}, val={len(val_samples)}, test={len(test_samples)}")
return DataSplit(
train=train_samples,
validation=val_samples,
test=test_samples,
split_info=split_info,
)
def create_k_folds(
self,
samples: list[DatasetSample],
k: int = 5,
shuffle: bool = True,
) -> list[CrossValidationFold]:
"""
Create k-fold cross-validation splits.
Args:
samples: List of all samples
k: Number of folds
shuffle: Whether to shuffle before splitting
Returns:
List of CrossValidationFold objects
"""
if k < 2:
raise ValueError("k must be at least 2")
if len(samples) < k:
raise ValueError(f"Need at least {k} samples for {k}-fold CV")
# Copy and optionally shuffle
all_samples = list(samples)
if shuffle:
self.rng.shuffle(all_samples)
# Calculate fold sizes
fold_size = len(all_samples) // k
folds = []
for fold_id in range(k):
# Validation is the current fold
val_start = fold_id * fold_size
val_end = len(all_samples) if fold_id == k - 1 else val_start + fold_size # noqa: SIM108
val_samples = all_samples[val_start:val_end]
train_samples = all_samples[:val_start] + all_samples[val_end:]
folds.append(
CrossValidationFold(
fold_id=fold_id,
train=train_samples,
validation=val_samples,
)
)
logger.info(f"Created {k}-fold cross-validation splits")
return folds
class StratifiedSplitter(DataSplitter):
"""
Stratified dataset splitter.
Ensures proportional representation of categories
(domain, difficulty, etc.) across splits.
"""
def __init__(self, seed: int = 42, stratify_by: str = "domain"):
"""
Initialize stratified splitter.
Args:
seed: Random seed for reproducibility
stratify_by: Attribute to stratify on ('domain', 'difficulty', 'labels')
"""
super().__init__(seed)
self.stratify_by = stratify_by
def split(
self,
samples: list[DatasetSample],
train_ratio: float = 0.7,
val_ratio: float = 0.15,
test_ratio: float = 0.15,
shuffle: bool = True,
) -> DataSplit:
"""
Stratified split maintaining category proportions.
Args:
samples: List of all samples
train_ratio: Proportion for training
val_ratio: Proportion for validation
test_ratio: Proportion for testing
shuffle: Whether to shuffle before splitting
Returns:
DataSplit with stratified train, validation, and test sets
"""
if abs(train_ratio + val_ratio + test_ratio - 1.0) > 0.001:
raise ValueError("Ratios must sum to 1.0")
if not samples:
raise ValueError("Cannot split empty sample list")
# Group samples by stratification key
groups = defaultdict(list)
for sample in samples:
key = self._get_stratify_key(sample)
groups[key].append(sample)
# Split each group proportionally
train_samples = []
val_samples = []
test_samples = []
for _key, group_samples in groups.items():
if shuffle:
self.rng.shuffle(group_samples)
n = len(group_samples)
train_end = int(n * train_ratio)
val_end = train_end + int(n * val_ratio)
train_samples.extend(group_samples[:train_end])
val_samples.extend(group_samples[train_end:val_end])
test_samples.extend(group_samples[val_end:])
# Final shuffle of combined sets
if shuffle:
self.rng.shuffle(train_samples)
self.rng.shuffle(val_samples)
self.rng.shuffle(test_samples)
# Verify stratification
stratify_info = self._verify_stratification(train_samples, val_samples, test_samples)
split_info = {
"total_samples": len(samples),
"train_samples": len(train_samples),
"val_samples": len(val_samples),
"test_samples": len(test_samples),
"train_ratio": len(train_samples) / len(samples),
"val_ratio": len(val_samples) / len(samples),
"test_ratio": len(test_samples) / len(samples),
"stratify_by": self.stratify_by,
"stratification_info": stratify_info,
"seed": self.seed,
"shuffled": shuffle,
}
logger.info(
f"Stratified split ({self.stratify_by}): "
f"train={len(train_samples)}, val={len(val_samples)}, "
f"test={len(test_samples)}"
)
return DataSplit(
train=train_samples,
validation=val_samples,
test=test_samples,
split_info=split_info,
)
def _get_stratify_key(self, sample: DatasetSample) -> str:
"""Get stratification key for a sample."""
if self.stratify_by == "domain":
return sample.domain or "unknown"
elif self.stratify_by == "difficulty":
return sample.difficulty or "unknown"
elif self.stratify_by == "labels":
return ",".join(sorted(sample.labels)) if sample.labels else "unknown"
else:
return str(getattr(sample, self.stratify_by, "unknown"))
def _verify_stratification(
self,
train: list[DatasetSample],
val: list[DatasetSample],
test: list[DatasetSample],
) -> dict[str, dict[str, float]]:
"""
Verify that stratification was successful.
Returns dictionary showing distribution of stratification key
across train/val/test splits.
"""
def get_distribution(samples: list[DatasetSample]) -> dict[str, float]:
if not samples:
return {}
counts = defaultdict(int)
for sample in samples:
key = self._get_stratify_key(sample)
counts[key] += 1
total = len(samples)
return {k: v / total for k, v in counts.items()}
return {
"train": get_distribution(train),
"validation": get_distribution(val),
"test": get_distribution(test),
}
def create_stratified_k_folds(
self,
samples: list[DatasetSample],
k: int = 5,
shuffle: bool = True,
) -> list[CrossValidationFold]:
"""
Create stratified k-fold cross-validation splits.
Args:
samples: List of all samples
k: Number of folds
shuffle: Whether to shuffle before splitting
Returns:
List of CrossValidationFold objects with stratification
"""
if k < 2:
raise ValueError("k must be at least 2")
# Group samples by stratification key
groups = defaultdict(list)
for sample in samples:
key = self._get_stratify_key(sample)
groups[key].append(sample)
# Initialize folds
folds_data = [{"train": [], "val": []} for _ in range(k)]
# Distribute each group across folds
for _key, group_samples in groups.items():
if shuffle:
self.rng.shuffle(group_samples)
# Assign samples to folds
fold_size = len(group_samples) // k
for fold_id in range(k):
val_start = fold_id * fold_size
val_end = len(group_samples) if fold_id == k - 1 else val_start + fold_size
for i, sample in enumerate(group_samples):
if val_start <= i < val_end:
folds_data[fold_id]["val"].append(sample)
else:
folds_data[fold_id]["train"].append(sample)
# Create fold objects
folds = [
CrossValidationFold(
fold_id=i,
train=data["train"],
validation=data["val"],
)
for i, data in enumerate(folds_data)
]
logger.info(f"Created stratified {k}-fold cross-validation splits")
return folds
class BalancedSampler:
"""
Balanced sampling for imbalanced datasets.
Provides utilities for:
- Oversampling minority classes
- Undersampling majority classes
- SMOTE-like synthetic sampling (for numerical features)
"""
def __init__(self, seed: int = 42):
"""Initialize balanced sampler."""
self.seed = seed
import random
self.rng = random.Random(seed)
def oversample_minority(
self,
samples: list[DatasetSample],
target_key: str = "domain",
target_ratio: float = 1.0,
) -> list[DatasetSample]:
"""
Oversample minority classes to balance dataset.
Args:
samples: Original samples
target_key: Attribute to balance on
target_ratio: Target ratio relative to majority (1.0 = equal)
Returns:
Balanced sample list (originals + oversampled)
"""
# Group by target key
groups = defaultdict(list)
for sample in samples:
key = getattr(sample, target_key, "unknown") or "unknown"
groups[key].append(sample)
# Find majority class size
max_count = max(len(g) for g in groups.values())
target_count = int(max_count * target_ratio)
# Oversample minority classes
balanced = []
for _key, group in groups.items():
balanced.extend(group)
# Oversample if needed
if len(group) < target_count:
num_to_add = target_count - len(group)
for _ in range(num_to_add):
# Randomly duplicate from group
original = self.rng.choice(group)
duplicate = DatasetSample(
id=f"{original.id}_oversample_{self.rng.randint(0, 999999)}",
text=original.text,
metadata={**original.metadata, "oversampled": True},
labels=original.labels,
difficulty=original.difficulty,
domain=original.domain,
reasoning_steps=original.reasoning_steps,
)
balanced.append(duplicate)
logger.info(f"Oversampled from {len(samples)} to {len(balanced)} samples")
return balanced
def undersample_majority(
self,
samples: list[DatasetSample],
target_key: str = "domain",
target_ratio: float = 1.0,
) -> list[DatasetSample]:
"""
Undersample majority classes to balance dataset.
Args:
samples: Original samples
target_key: Attribute to balance on
target_ratio: Target ratio relative to minority (1.0 = equal)
Returns:
Balanced sample list (subset of originals)
"""
# Group by target key
groups = defaultdict(list)
for sample in samples:
key = getattr(sample, target_key, "unknown") or "unknown"
groups[key].append(sample)
# Find minority class size
min_count = min(len(g) for g in groups.values())
target_count = int(min_count * target_ratio)
# Undersample majority classes
balanced = []
for _key, group in groups.items():
if len(group) > target_count:
# Randomly select target_count samples
balanced.extend(self.rng.sample(group, target_count))
else:
balanced.extend(group)
logger.info(f"Undersampled from {len(samples)} to {len(balanced)} samples")
return balanced
def get_class_distribution(
self,
samples: list[DatasetSample],
target_key: str = "domain",
) -> dict[str, int]:
"""
Get distribution of classes.
Args:
samples: Sample list
target_key: Attribute to analyze
Returns:
Dictionary of class counts
"""
distribution = defaultdict(int)
for sample in samples:
key = getattr(sample, target_key, "unknown") or "unknown"
distribution[key] += 1
return dict(distribution)
|