Training Data
Managing and improving training data quality.
Data Labeling Strategies
Manual Labeling
# Export for Label Studio
def export_for_labeling(data: pd.DataFrame, output_path: str):
tasks = [
{"data": {"text": row["text"]}, "id": idx}
for idx, row in data.iterrows()
]
with open(output_path, 'w') as f:
json.dump(tasks, f)
Weak Supervision
from snorkel.labeling import labeling_function, LabelingFunction
from snorkel.labeling.model import LabelModel
@labeling_function()
def lf_keyword(x):
keywords = ["urgent", "free", "winner"]
return 1 if any(k in x.text.lower() for k in keywords) else -1
@labeling_function()
def lf_short_text(x):
return 1 if len(x.text) < 20 else -1
# Combine weak labels
lfs = [lf_keyword, lf_short_text]
applier = PandasLFApplier(lfs)
L_train = applier.apply(df_train)
label_model = LabelModel(cardinality=2)
label_model.fit(L_train, n_epochs=100)
labels = label_model.predict(L_train)
Active Learning
from modAL.models import ActiveLearner
from sklearn.ensemble import RandomForestClassifier
learner = ActiveLearner(
estimator=RandomForestClassifier(),
query_strategy=uncertainty_sampling,
X_training=X_initial, y_training=y_initial
)
for _ in range(n_queries):
query_idx, query_instance = learner.query(X_pool)
# Human labels the instance
y_new = get_human_label(query_instance)
learner.teach(X_pool[query_idx], y_new)
X_pool = np.delete(X_pool, query_idx, axis=0)
Data Augmentation
# Text augmentation
import nlpaug.augmenter.word as naw
aug = naw.SynonymAug(aug_src='wordnet')
augmented = aug.augment("The quick brown fox jumps over the lazy dog")
# Image augmentation
import albumentations as A
transform = A.Compose([
A.RandomCrop(width=256, height=256),
A.HorizontalFlip(p=0.5),
A.RandomBrightnessContrast(p=0.2),
A.Normalize()
])
# Tabular augmentation (SMOTE)
from imblearn.over_sampling import SMOTE
smote = SMOTE(random_state=42)
X_resampled, y_resampled = smote.fit_resample(X_train, y_train)
Handling Imbalanced Data
# Class weights
from sklearn.utils.class_weight import compute_class_weight
class_weights = compute_class_weight('balanced', classes=np.unique(y), y=y)
# Focal loss
def focal_loss(y_true, y_pred, gamma=2, alpha=0.25):
bce = F.binary_cross_entropy(y_pred, y_true, reduction='none')
p_t = y_true * y_pred + (1 - y_true) * (1 - y_pred)
focal_weight = (1 - p_t) ** gamma
return (alpha * focal_weight * bce).mean()
# Stratified sampling
from sklearn.model_selection import StratifiedKFold
skf = StratifiedKFold(n_splits=5, shuffle=True)
for train_idx, val_idx in skf.split(X, y):
X_train, X_val = X[train_idx], X[val_idx]
Data Splitting
from sklearn.model_selection import train_test_split
# Random split (with stratification)
X_train, X_test, y_train, y_test = train_test_split(
X, y, test_size=0.2, stratify=y, random_state=42
)
# Temporal split (for time series)
def temporal_split(df, time_col, train_end, val_end):
train = df[df[time_col] < train_end]
val = df[(df[time_col] >= train_end) & (df[time_col] < val_end)]
test = df[df[time_col] >= val_end]
return train, val, test
# Group split (no data leakage)
from sklearn.model_selection import GroupShuffleSplit
gss = GroupShuffleSplit(n_splits=1, test_size=0.2)
for train_idx, test_idx in gss.split(X, y, groups=user_ids):
X_train, X_test = X[train_idx], X[test_idx]
Commands
/omgdata:label- Data labeling/omgdata:augment- Augmentation/omgdata:split- Data splitting
Best Practices
- Start with a small, high-quality labeled set
- Use weak supervision to scale labeling
- Match augmentation to your domain
- Prevent data leakage in splits
- Monitor label quality over time
Converted and distributed by TomeVault — claim your Tome and manage your conversions.