Key Terms

Important terminology and concepts used in Tracebloc.

​

Platform Terms

​

Data owner

The person or organization that deploys the secure environment, ingests datasets, creates use cases, and decides who can join them.

​

Data scientist

An invited peer who uploads models and runs experiments on a use case. Data scientists never see the raw data.

​

Use Case

A specific AI project or evaluation scenario where data owners and data scientists collaborate.

​

Secure environment

The tracebloc software deployed on the data owner’s infrastructure that stores datasets and runs model training. Older docs and some Kubernetes resource names call it the client.

​

Dataset

Training or test data ingested into a secure environment. Each dataset has an intent: train for training data or test for the held-out data that scores models.

​

Task type

The machine learning task a dataset and its models address, for example image classification or object detection.

​

Model

An AI/ML model submitted by data scientists for evaluation on the data owner’s infrastructure.

​

Model zoo

The tracebloc model zoo: ready-made model templates for every task type.

​

Seed weights

The pretrained weights of a model zoo template. Fetch them with user.fetch_seed_weights(); without them, the template trains from random weights.

​

Experiment

One attempt by a data scientist to train a model on a use case’s dataset.

​

Cycle

One round of federated training: each secure environment trains the model locally, then tracebloc combines the weight updates.

​

Technical Terms

​

Federated Learning

A machine learning approach where models are trained across multiple devices while keeping data local.

​

Differential Privacy

A system for publicly sharing information about a dataset while withholding information about individuals. tracebloc does not apply differential privacy to experiments by default, and the dashboard has no setting to turn it on.

​

Model Weights

The parameters of a trained model that determine its behavior and performance.

​

Process Terms

​

Training

The process of teaching a model to make predictions based on data.

​

Evaluation

The process of assessing a model’s performance using predefined metrics.

​

Fine-tuning

The process of adjusting a pre-trained model for a specific task.

​

Next Steps

• Check FAQs
• Review SDK documentation