Dataset Curator

What it does

The Dataset Curator skill assists users in preparing high-quality training data for machine learning models. It guides you through a systematic process of cleaning, validating, augmenting, and maintaining datasets to improve model performance. The skill covers tasks like deduplication, label quality assessment, bias detection, augmentation strategies, and version control, applying best practices from production ML systems.

When to use it

• When building classifiers or fine-tuning large language models (LLMs).
• When training custom machine learning models where data quality is critical.
• For ongoing maintenance of existing datasets used for model training.
• To identify and address issues like duplicates, mislabeled examples, outliers, and bias in your dataset.
• When needing to balance class distributions within a dataset.

Key capabilities

• Dataset Profiling: Analyzes size, dimensionality, label distribution, missing values, and feature statistics.
• Duplicate Detection: Identifies exact and near-duplicate entries using hash-based and similarity-based methods.
• Label Quality Assessment: Identifies likely mislabeled examples and provides confidence scoring for review.
• Bias Detection: Detects bias and representation gaps within the dataset.
• Data Augmentation: Generates new data points through techniques like paraphrasing, image transformations (rotation, flips), and feature perturbation.
• Class Balancing: Applies oversampling (SMOTE, random) and undersampling strategies to address class imbalance.

Example prompts

• "Check quality of this dataset."
• "Find duplicates in dataset."
• "Assess label balance in my training data."

Tips & gotchas

• Data quality is a critical factor for model performance, so prioritize using this skill early in your ML workflow.
• The skill helps understand missingness mechanisms (MCAR, MAR, MNAR) to appropriately handle missing values.
• When augmenting data, validate the generated samples to ensure they are realistic and don't introduce new biases.