Entity Concepts

Entities are objects that you create and manage while using the NeMo microservices. They include projects, namespaces, datasets, models, job configurations, job information, job results, and more. Some entities are specific to microservices, while others are shared across multiple microservices.

Organizational Entities

Organizational entities help structure and manage resources in the NeMo microservices platform. They provide a hierarchical organization system that enables efficient management of datasets, models, and other assets.

Namespaces

Namespaces are for organizing all other entities into logical groups. They represent the highest level of organization in NeMo Entity Store. You can use namespaces to organize your entities by team, user, product, or some other criteria.

Projects

Projects serve as organizational containers that group related entities such as datasets, models, customized models, and evaluation results. In the NeMo Entity Store hierarchy, projects represent the second-highest level of organization, allowing you to structure and manage related resources for specific objectives.

Entities with Files

While the NeMo Entity Store microservice manages the entities, the NeMo Data Store microservice handles the actual files associated with those entities, if any, and supports the Hugging Face Hub (HfApi) interface. For example, a dataset entity can be associated with a set of files such as train.json, validation.json, and test.json. The NeMo Data Store microservice is responsible for managing the files associated with these entities, including uploading, downloading, and versioning.

Datasets

Datasets contain files you use in model fine-tuning and evaluation. Depending on your use case, to use the dataset files with NeMo Customizer and NeMo Evaluator, you might need to make sure the data is in a specific format. For more guidance, check out the following pages:

• Format Training Dataset

• Custom Evaluations

• Using Custom Data

Models

The following are several kinds of models you can manage in the NeMo microservices platform:

• Base model: A base model is a pre-trained model, which is a starting point for training or fine-tuning, typically in .nemo format. You can also use these models for inference.

• Fine-tuned model: A fine-tuned model is a model trained on an additional dataset specific to your use case, on top of a base model.

• Full SFT model: A full supervised fine-tuning (SFT) model is a model that has undergone comprehensive fine-tuning on a specific dataset, typically resulting in more significant modifications to the base model’s behavior.

• Guardrails model: A guardrails model is a model associated with a guardrails configuration defined in NeMo Guardrails. Guardrails are used to enforce safety, security, and compliance boundaries around AI model interactions.

• Prompt-engineered model: A prompt-engineered model is a model that has been optimized through systematic prompt design and testing, without modifying the underlying model weights.

• API endpoint model: An API endpoint model is a model that has been deployed and exposed through an API interface, allowing for programmatic access to model inference capabilities.