Getting Started with MLflow: MLflow Dataset Tracking

This episode shows how MLflow’s built-in dataset tracking lets you version, log, and manage datasets just like models — no extra tools required. The big focus is reproducibility: knowing exactly what data was used in every experiment so results can be recreated without guesswork. It walks through real ML workflows, like handling massive datasets (e.g., Wikipedia-scale text), refining subsets with LLMs, and keeping data consistent across multiple experiment runs. The video explains how MLflow can: 🔹 Log datasets from tools like pandas, Hugging Face, and Spark 🔹 Automatically capture dataset schemas 🔹 Track version history as data evolves 🔹 Use APIs (like dataset creation tools) for annotations and evaluation data 🔹 Combine outputs from multiple runs into structured datasets The core idea: data changes are just as important as model changes. Without tracking them, experiments become impossible to reproduce. MLflow’s dataset versioning creates a clear history of how data was selected, modified, and used, which helps teams avoid surprises, reduce memory issues, retrain models more efficiently, and collaborate better. 🔗 GitHub link: https://github.com/srinathmkce/TheAIGuy/blob/main/NLP/mlflow_examples/3-dataset-tracking.ipynb 🎤 Speaker: Shrinath Suresh, MLflow Ambassador 00:00 Understanding the Need for Dataset Tracking 01:59 Importance of Data Tracking and Implementation 02:05 Setting Up MLflow and Dataset Preparation 03:43 Keyword-Based Search and Filtering 04:32 Introduction to LLM Output and Article Selection 06:35 Data Versioning and Tracking in MLflow 07:07 Introduction to MLflow Datasets 08:59 Preparing Data for Summarization 09:20 Introduction to Evaluation Phase 10:29 Evaluation Results Overview 13:58 Searching and Annotating Datasets 14:49 The Importance of Reproducibility in Machine Learning