dataset-curator（数据集管理员）

Dataset Curator

The Dataset Curator skill guides you through the critical process of preparing high-quality training data for machine learning models. Data quality is the single most important factor in model performance, yet it is often underinvested. This skill helps you systematically clean, validate, augment, and maintain datasets that lead to better models.

From initial collection to ongoing maintenance, this skill covers deduplication, label quality assessment, bias detection, augmentation strategies, and version control. It applies best practices from production ML systems to ensure your datasets are not just clean, but strategically optimized for your learning objectives.

Whether you are building a classifier, fine-tuning an LLM, or training a custom model, this skill ensures your data foundation is solid.

Core Workflows

Workflow 1: Assess Dataset Quality

1. Profile the dataset:

- Size and dimensionality - Label distribution and balance - Missing value patterns - Feature statistics

1. Identify quality issues:

- Duplicates (exact and near-duplicate) - Mislabeled examples - Outliers and anomalies - Data leakage - Bias and representation gaps

1. Measure quality metrics: def assess_quality(dataset): return {"size": len(dataset), "duplicate_rate": find_duplicates(dataset).ratio, "missing_rate": dataset.isnull().mean(), "label_balance": compute_entropy(dataset.labels), "outlier_rate": detect_outliers(dataset).ratio, "estimated_label_noise": estimate_label_noise(dataset)}
2. Prioritize issues by impact
3. Create remediation plan

Workflow 2: Clean and Prepare Data

1. Remove duplicates:

- Exact duplicates: hash-based dedup - Near-duplicates: similarity-based clustering - Decide: keep first, best, or merge

1. Handle missing values:

- Understand missingness mechanism (MCAR, MAR, MNAR) - Impute, drop, or flag appropriately

1. Fix label quality:

- Identify likely mislabels with confidence scoring - Route to human review or automatic correction - Document labeling guidelines

1. Normalize and standardize:

- Consistent formatting - Schema validation - Encoding standardization

1. Validate cleaned dataset

Workflow 3: Augment and Balance

1. Analyze class imbalance:

- Compute imbalance ratios - Assess impact on model training

1. Apply balancing strategies:

- Oversampling minority classes (SMOTE, random) - Undersampling majority classes - Class weights in training

1. Generate augmentations:

- Text: paraphrase, synonym substitution, back-translation - Image: rotation, flip, color jitter, mixup - Tabular: noise injection, feature perturbation

1. Validate augmentation quality:

- Ensure augmented samples are realistic - Check for introduced biases

1. Version and document changes

Quick Reference

Best Practices

• Profile Before Processing: Understand your data before changing it

- Compute statistics and visualize distributions - Document original state for reference - Identify patterns in issues

• Preserve Provenance: Track every transformation

- Version control datasets like code - Log all cleaning operations - Maintain mapping between original and cleaned data

• Prioritize Label Quality: Garbage labels in, garbage model out

- Invest in clear labeling guidelines - Use multiple annotators and measure agreement - Regular quality audits of labels

• Test Cleaning Impact: Measure effect of cleaning

- Train models on original vs cleaned data - Track which cleaning steps help most - Avoid cleaning that hurts performance

• Stratify Splits Carefully: Maintain distribution in train/val/test

- Stratify by label and key features - Keep related samples in same split - Ensure temporal ordering if applicable

• Document Everything: Future you will thank present you

- Dataset cards with key statistics - Known issues and limitations - Collection methodology and biases

Advanced Techniques

Confident Learning for Label Noise

Identify and fix mislabeled examples:

from cleanlab import find_label_issues

# Train model to get predicted probabilities
model.fit(X_train, y_train)
pred_probs = model.predict_proba(X_train)

# Find likely mislabeled examples
issues = find_label_issues(
    labels=y_train,
    pred_probs=pred_probs,
    return_indices_ranked_by="self_confidence"
)

# Review and correct top issues
for idx in issues[:100]:
    review_and_correct(X_train[idx], y_train[idx])

Similarity-Based Deduplication

Remove near-duplicates using embeddings:

def deduplicate_semantic(texts, threshold=0.95):
    embeddings = embed(texts)
    clusters = cluster_by_similarity(embeddings, threshold)

    # Keep one representative per cluster
    deduplicated = []
    for cluster in clusters:
        representative = select_best(cluster)  # longest, most recent, etc.
        deduplicated.append(representative)

    return deduplicated

Active Learning for Efficient Labeling

Prioritize labeling effort:

def active_learning_loop(unlabeled_pool, labeled_set, budget):
    while len(labeled_set) < budget:
        # Train on current labeled data
        model.fit(labeled_set)

        # Score unlabeled by uncertainty
        uncertainties = model.uncertainty(unlabeled_pool)

        # Select most uncertain for labeling
        to_label = select_top_k(unlabeled_pool, uncertainties, k=10)
        labels = human_label(to_label)

        # Update sets
        labeled_set.add(to_label, labels)
        unlabeled_pool.remove(to_label)

    return labeled_set

Data Slice Analysis

Find problematic subgroups:

def find_weak_slices(model, data, features):
    # Evaluate on all slices
    slices = generate_slices(data, features)

    weak_slices = []
    for slice_name, slice_data in slices:
        performance = evaluate(model, slice_data)
        if performance < overall_performance - threshold:
            weak_slices.append({
                "slice": slice_name,
                "size": len(slice_data),
                "performance": performance
            })

    return sorted(weak_slices, key=lambda x: x["performance"])

Common Pitfalls to Avoid

• Cleaning test data the same way as training data (causes leakage)
• Over-aggressive deduplication that removes valid variations
• Imputing values without understanding the missingness mechanism
• Augmenting in ways that introduce unrealistic examples
• Ignoring class imbalance until model training fails
• Not versioning datasets, making experiments irreproducible
• Assuming more data is always better (quality > quantity)
• Failing to document data collection biases and limitations