How It Works
Application Components
Transformer Lab has two main components:
-
Transformer Lab App: a graphical client application.
-
Transformer Lab API: an LLM server that is wrapped with an API
Underlying ML Frameworks
Transformer Lab wraps several popular ML frameworks that do the heavy lifting under the scenes.
Huggingface's open libraries are used to fetch, store, and train models using the Huggingface Transformers, Hub, Accelerate, PEFT, and Datasets libraries.
For Inference, Transformer Lab uses FastChat
Transformer Lab's main plugin for Evaluation calls to Eleuther Lab's Language Model Evaluation Harness
Automatic Server Provisioning
Especially when working with cloud servers, it's common to set up a new server from scratch right before starting to work with it.
Transformer Lab has a feature to autoprovision a server. This makes it easier to get started because you can just download the Client Application and point it at a (cloud or local) server through SSH.