Dataset Curator

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
2. Identify quality issues:
   • Duplicates (exact and near-duplicate)
   • Mislabeled examples
   • Outliers and anomalies
   • Data leakage
   • Bias and representation gaps
3. 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)
       }
   

4. Prioritize issues by impact
5. 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
2. Handle missing values:
   • Understand missingness mechanism (MCAR, MAR, MNAR)
   • Impute, drop, or flag appropriately
3. Fix label quality:
   • Identify likely mislabels with confidence scoring
   • Route to human review or automatic correction
   • Document labeling guidelines
4. Normalize and standardize:
   • Consistent formatting
   • Schema validation
   • Encoding standardization
5. Validate cleaned dataset