Overview

Every ML benchmark you have used before works the same way: download the data, iterate locally, submit a predictions file. It is a familiar loop — and it only works when the data is public. The moment it belongs to a hospital, a bank, or a manufacturer, that loop breaks entirely. This is your hands-on introduction to the alternative: a federated learning workspace where you submit code, not data, the data never moves, and your score comes back to you. A year-round binary classification challenge with no deadline and no pressure — designed so you can learn the mechanic before the stakes are real.

The Difference from Every Other ML Benchmark

Most ML practitioners have built intuition on Kaggle or similar platforms. The format is well understood: a dataset arrives as a download, you iterate in a local notebook, you submit a CSV of predictions, and you watch your leaderboard position. That workflow is excellent for public data. It is incompatible with the data that makes the most interesting ML problems — clinical records, financial transactions, industrial sensor logs, proprietary imagery. Organisations holding that data cannot hand it to external practitioners. Legal obligations, competitive sensitivity, and contractual constraints make it impossible.

The federated mechanic inverts the relationship between data and computation. Instead of data moving to your model, your model moves to the data. You write a training script, package it to the workspace specification, and submit it. That code executes inside the data owner's infrastructure, directly on the dataset. What returns to you is your evaluation score. The raw records never leave the environment — not to tracebloc's servers, not to your machine, not anywhere. This is distributed machine learning as it actually runs in production, not a simulated version of it using a public proxy dataset.

Understanding this mechanic is the prerequisite for every real enterprise challenge on the workspace — healthcare, finance, manufacturing, aviation. This introduction is where you build that understanding with a purpose-built starter dataset and no pressure attached.

The Starter Challenge

The introduction challenge is a binary classification task that runs year-round. There is no submission deadline, no prize pool, and no competitive pressure. It exists as a federated learning sandbox — a place to learn the submission format, understand the evaluation pipeline, and build the iteration habits that carry over directly to live enterprise challenges.

Practitioners who work through this tutorial first consistently find the transition to real enterprise challenges smoother, both technically and conceptually. The first few submissions to a live challenge should be spent improving your model, not debugging submission format errors. This is where you eliminate that cost before it matters.

Think of it as the tracebloc equivalent of a development environment: work here until the workflow is second nature, then step into challenges where the data is real and the impact is real too.

The Dataset

Binary Classification — Intro Dataset

The intro dataset is structured tabular data with a binary target label. It is intentionally approachable: clean feature columns, no exotic preprocessing requirements, and a target distribution that makes standard classification approaches viable from the first submission. The goal is not to make the ML problem hard — it is to make the federated submission mechanics easy to learn through a concrete, hands-on federated learning example.

Because the dataset lives inside the workspace, there is no download link. Instead, explore it through the EDA tab, which surfaces summary statistics, feature distributions, and correlation views computed directly on the data. The EDA tab is itself a demonstration of the core principle: you get meaningful insight into the data without it ever leaving the workspace. Spend time here before writing your first training script — understanding the feature space and target balance early saves significant iteration time later.

How the Workspace Works

The Submission Model

The submission flow is different enough from prediction-based benchmarks that it is worth walking through explicitly before your first attempt.

Study the data. Open the EDA tab and understand the feature schema, target distribution, and any class imbalance. Your script will need to handle this data correctly without you ever seeing a raw row.
Write your training script. Build a model locally using the schema published in the workspace description. Your script should load data, train a model, and output predictions or evaluation metrics in the format the pipeline expects.
Package and submit. tracebloc expects code in a defined format — a Python script or module conforming to the workspace interface specification. Submit through the training and fine-tuning tab.
Your code runs on the private dataset. Inside the infrastructure, your code executes against the actual dataset. At no point is the data transmitted to you or to any external system. This is ML without uploading data, as it operates in production.
Your score is returned. The evaluation result is computed inside the secure environment and surfaced as a score. The leaderboard is updated.
Iterate. Adjust your model, resubmit, and track your progress on the leaderboard. The challenge runs year-round — there is no cost to a slow start.

This is a federated learning application in practice, not an approximation of one. The constraint — submit code, not data — shapes how you write code, how you debug, and how you think about what information you actually need in order to build a good model.

What You Learn Here

Working through this tutorial builds a specific set of skills that transfer directly to real enterprise challenges.

You learn to write code that runs remotely on data you cannot inspect directly — which demands more careful thinking about schema assumptions, edge cases, and error handling than local iteration typically requires. You get familiar with the evaluation pipeline: what it measures, how scores are computed, and how to read the output. You practice using the EDA tab as your primary window into the dataset rather than a local notebook. And you develop intuition for iterating on model performance when your feedback loop runs through a remote execution environment rather than an interactive REPL.

These are not tracebloc-specific curiosities. They are the practical skills required for any privacy preserving AI application where the data cannot move — which covers the majority of the most valuable ML problems in healthcare, financial services, and industrial operations. The federated learning vs centralised distinction is not academic once you are working on data that genuinely cannot leave an institution.

Getting Started

Open the EDA tab first. Understand the feature distributions and target balance before writing a single line of model code. Read the evaluation metric in the description tab — knowing what you are optimising for from the start saves significant iteration later.

When you are ready, follow the submission format specification in the description tab and submit your first model through the training and fine-tuning tab. Check your score and review your position on the leaderboard. Then iterate. The challenge runs continuously — there is no deadline and no pressure to rush the fundamentals.

Disclaimer

The dataset used in this introduction challenge is a purpose-built starter dataset. It is designed to be representative of the tabular binary classification problems that appear in real enterprise challenges, but is not itself real enterprise sensitive data. The federated execution mechanic — your code runs on the data inside the workspace, data never leaves — operates identically to how it works in live challenges with healthcare, finance, or manufacturing datasets. What differs is the stakes, not the mechanics.

Federated Learning Tutorial: Run Your First Model on Private Data

Federated Learning Tutorial: Run Your First Model on Private Data